Video cold start method and device, storage medium and electronic equipment

Abstract

The invention discloses a video cold start method and device, a storage medium and electronic equipment, and belongs to the technical field of artificial intelligence, and the method comprises the steps: obtaining a target video which is obtained at a target moment; inputting the target video into the cold start behavior prediction model to obtain a cold start behavior prediction value, the cold start behavior prediction value indicating the probability of triggering the target behavior after the target video is cold started; determining cold start test flow according to the cold start behavior prediction value; performing cold start on the target video based on the cold start test flow; training samples of the cold start behavior prediction model are constructed based on a sample video published in a target time period and a target behavior generated by triggering the sample video in the target time period, the length of the target time period is a preset length, and a difference value between an ending moment of the target time period and a target moment is smaller than the preset length. According to the method, the cold start behavior prediction model is specially trained for cold start, so that the accurate heuristic flow is determined for the prediction result of the new video, and the cold start effect is remarkably improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to video cold start methods, apparatus, storage media, and electronic devices. Background Technology

[0002] Video recommendation services in related technologies face the video cold start problem. The video cold start problem refers to the fact that for a new video, the video recommendation service lacks sufficient user feedback data for the new video, as well as sufficient modeling information for the new video itself. As a result, the video recommendation service has difficulty determining how many users to recommend the new video to and which users to recommend it to.

[0003] To address the video cold start problem, video recommendation services preset trial traffic and recommend new videos within the trial traffic range to achieve a cold start for new videos. However, this solution suffers from poor video cold start performance because it cannot accurately determine the trial traffic. Summary of the Invention

[0004] This application provides a video cold start method, apparatus, storage medium, and electronic device, which can significantly improve the video cold start effect by training a cold start behavior prediction model specifically for video cold starts and determining accurate trial traffic based on the prediction results of the cold start behavior prediction model for new videos.

[0005] According to one aspect of the embodiments of this application, a video cold start method is provided, the method comprising:

[0006] Acquire the target video, which is acquired at a target time;

[0007] The target video is input into the cold start behavior prediction model to obtain the cold start behavior prediction value, which indicates the probability that the target video will trigger the target behavior after being cold-started.

[0008] Based on the predicted cold start behavior, determine the cold start trial flow;

[0009] Based on the cold start probe traffic, a cold start is performed on the target video;

[0010] The training samples of the cold start behavior prediction model are all constructed based on sample videos released within the target time period and the target behaviors triggered by the sample videos during the target time period. The length of the target time period is a preset length, and the difference between the end time of the target time period and the target time is less than the preset length.

[0011] According to one aspect of the embodiments of this application, a video cold start device is provided, the device comprising:

[0012] A video acquisition module is used to acquire a target video, which is acquired at a target time.

[0013] The cold start module is used to perform the following operations:

[0014] The target video is input into the cold start behavior prediction model to obtain the cold start behavior prediction value, which indicates the probability that the target video will trigger the target behavior after being cold-started.

[0015] Based on the predicted cold start behavior, determine the cold start trial flow;

[0016] Based on the cold start probe traffic, a cold start is performed on the target video;

[0017] The training samples of the cold start behavior prediction model are all constructed based on sample videos released within the target time period and the target behaviors triggered by the sample videos during the target time period. The length of the target time period is a preset length, and the difference between the end time of the target time period and the target time is less than the preset length.

[0018] According to one aspect of the embodiments of this application, a computer device is provided, the computer device including a processor and a memory, the memory storing at least one instruction, at least one program, code set or instruction set, the at least one instruction, the at least one program, the code set or instruction set being loaded and executed by the processor to implement the above-described video cold start method.

[0019] According to one aspect of the embodiments of this application, a computer-readable storage medium is provided, wherein the storage medium stores at least one instruction, at least one program, code set or instruction set, wherein the at least one instruction, the at least one program, the code set or instruction set is loaded and executed by a processor to implement the above-described video cold start method.

[0020] According to one aspect of this application, a computer program product is provided, the computer program product including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, causing the computer device to perform the above-described video cold start method.

[0021] The technical solution provided in this application can bring the following beneficial effects:

[0022] This application provides a video cold start method, apparatus, storage medium, and electronic device. The embodiments of this application specifically train a cold start behavior prediction model for video cold starts, without reusing recommendation models or solely considering video quality. Compared to related technologies, it can more accurately predict various user-related target behaviors that may be triggered after a target video (new video) cold starts. Based on this prediction, it determines appropriate cold start probe traffic for the target video, thereby improving the cold start effect.

[0023] This cold start behavior prediction model is trained using training samples from a short time period adjacent to the target video (new video), thus distinguishing it from recommendation models in related technologies. The training samples for the cold start behavior prediction model are from a short period, therefore the number of samples is not too large, avoiding prediction inaccuracies caused by too many consumer-side samples. Furthermore, the target time period corresponding to the samples is highly close to the time when the new video is acquired, thus avoiding prediction inaccuracies caused by long time intervals. This allows the cold start behavior prediction model to reflect the latest interest and behavioral tendencies of the user group in new videos.

[0024] This cold start behavior prediction model only serves cold start scenarios and is not used in conjunction with video recommendation in non-cold start scenarios, nor is it used for recommending videos other than new videos. This ensures the purity of the cold start behavior prediction model and avoids performance degradation caused by mixed use. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a schematic diagram of an application runtime environment provided in one embodiment of this application;

[0027] Figure 2 This is a flowchart of a video cold start method provided in one embodiment of this application;

[0028] Figure 3 This is a schematic flowchart of a cold start trial flow determination method provided in one embodiment of this application;

[0029] Figure 4 This is a schematic diagram of the cold start behavior prediction model structure provided in one embodiment of this application. Figure 1 ;

[0030] Figure 5 This is a schematic diagram of the cold start behavior prediction model structure provided in one embodiment of this application. Figure 2 ;

[0031] Figure 6 This is a schematic diagram of the training process of a cold start behavior prediction model provided in one embodiment of this application;

[0032] Figure 7 This is a schematic diagram of the entire cold start process provided in one embodiment of this application;

[0033] Figure 8 This is a schematic diagram illustrating an application scenario of a cold start solution provided in one embodiment of this application;

[0034] Figure 9 This is a block diagram of a video cold start device provided in one embodiment of this application;

[0035] Figure 10 This is a structural block diagram of a computer device provided in one embodiment of this application. Detailed Implementation

[0036] Before describing the specific embodiments of this application, the relevant technical background related to the embodiments of this application will be introduced to facilitate understanding by those skilled in the art.

[0037] Video recommendation services in related technologies face the video cold start problem. The video cold start problem refers to the fact that for new videos, the video recommendation service lacks sufficient user feedback data for the new video, as well as sufficient modeling information for the new video itself, such as rating information, classification information, etc., which makes it difficult for the video recommendation service to determine how many users to recommend the new video to and which users to recommend it to.

[0038] To address the cold start problem for videos, video recommendation services typically employ a tiered traffic guarantee scheme. This scheme is divided into a minimum guarantee phase and a support phase. The minimum guarantee phase aims to provide new videos with a trial traffic, such as 200 views, to gather user feedback data within a very limited traffic range. Views (vv) are typically used to measure the number of user visits; 200vv means 200 visits, i.e., the new video is pushed to 200 users to obtain 200 views of trial traffic. The support phase builds upon the user feedback data obtained in the minimum guarantee phase, selecting appropriate support traffic based on this feedback to further promote the new video. This approach avoids wasting traffic while effectively promoting new videos.

[0039] In some technical solutions, to determine reasonable trial traffic for the minimum wage stage, new videos can be processed based on content understanding. The video quality is then determined based on the processing results, and the trial traffic is determined based on the video quality. However, video quality is not directly related to the user feedback data generated by the new video. Although in some cases, there may be a positive correlation between video quality and the user feedback data generated by the new video, this positive correlation is not always valid. Therefore, solutions that determine trial traffic based on video quality cannot obtain reasonable trial traffic. Furthermore, these solutions do not consider user behavior, only the quality of the new video itself, which also leads to the determination of an unreasonable trial traffic.

[0040] Other solutions incorporate the new video into the recommendation model, determining trial traffic based on the model's output. The recommendation model is the foundational model used by video recommendation services. Its key characteristic is its ability to output video recommendation scores through extensive training on a large number of samples over a long period. For example, a high recommendation score indicates the video is worth recommending, or that it will likely receive positive user feedback after exposure. These solutions leverage this feature of recommendation models, inputting new videos and selecting higher trial traffic if the model outputs a high score, and vice versa. However, these solutions are limited by the recommendation model itself. Because the model is trained using a large number of samples from video consumption scenarios, it prioritizes consumption performance over the video's content. This limitation becomes more pronounced with larger sample sizes. New videos lack consumer engagement, causing recommendation models to output inaccurate results due to insufficient content. This leads to inaccurate trial traffic obtained from these videos. Furthermore, using recommendation models to process new videos, while addressing cold starts, negatively impacts their performance in non-cold-start scenarios, such as reducing the accuracy of recommendation predictions. In short, video recommendation service models are not suitable for direct use in cold starts of new videos, nor are they suitable for trial traffic used to determine video cold starts.

[0041] In view of this, embodiments of this application provide a video cold start method, which is applied to video cold start scenarios. This video cold start method can significantly improve the video cold start effect by training a cold start behavior prediction model specifically for video cold starts and determining accurate trial traffic based on the prediction results of the cold start behavior prediction model for new videos.

[0042] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings. It should be noted that all data used in the embodiments of this application has been fully authorized by the relevant parties before use.

[0043] Please refer to Figure 1 This diagram illustrates an application runtime environment provided in one embodiment of this application under an exemplary implementation. The application runtime environment may include: terminal 10 and server 20.

[0044] Terminal 10 includes, but is not limited to, electronic devices such as mobile phones, computers, smart voice interaction devices, smart home appliances, in-vehicle terminals, game consoles, e-book readers, multimedia playback devices, and wearable devices. Application clients can be installed on terminal 10.

[0045] In this embodiment, the application described above can be any application that runs a video recommendation service and supports video cold start. Typically, the application is a video application. Of course, in addition to video applications, other types of applications can also run video recommendation services and support video cold start. For example, news applications, social applications, search applications, interactive entertainment applications, browser applications, shopping applications, content sharing applications, virtual reality (VR) applications, augmented reality (AR) applications, etc., are not limited in this embodiment. Optionally, the terminal 10 runs a client of the above-mentioned application.

[0046] Server 20 provides background services to clients of applications in terminal 10. For example, server 20 can be a background server for the aforementioned applications. Server 20 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms. Optionally, server 20 can simultaneously provide background services to applications in multiple terminals 10.

[0047] Optionally, terminal 10 and server 20 can communicate with each other via network 30. Terminal 10 and server 20 can be directly or indirectly connected via wired or wireless communication, which is not limited herein.

[0048] Please refer to Figure 2 The document illustrates a flowchart of a video cold start method according to an exemplary embodiment. This method can be applied to computer devices, which refer to electronic devices capable of data computation and processing. For example, the entity executing each step may be... Figure 1 The relevant devices in the application's runtime environment are shown. This method may include the following steps:

[0049] Step S201. Acquire the target video, which is acquired at the target time.

[0050] In this embodiment, the target video is a new video obtained by the application. The application has not previously recommended the target video, lacks user feedback data for the target video, and lacks sufficient modeling information for the target video itself, such as rating information, classification information, etc.

[0051] This application does not limit the target video; it can be a short video or a medium-length video. With short video recommendations occupying an increasingly larger portion of people's lives, users can obtain desired content such as entertainment and knowledge by watching short videos. Medium-length videos, on the other hand, refer to videos that are longer, generally over one minute, requiring more effort to produce. The recommendation format also typically differs from the immersive flow of short videos, usually being a point-and-click approach. Both short video applications and medium-length video applications can be used to execute the video cold start method of this application.

[0052] Taking medium-length video applications as an example, they can recommend videos longer than 1 minute or even 5 minutes to users in a point-and-click manner. The production cost of medium-length videos is higher, thus requiring more effective cold start performance. For medium-length video applications, the user base is still relatively small. If related technologies are used to indiscriminately set trial traffic for each new medium-length video, the user base of the application cannot support it. Furthermore, many newly published medium-length videos do not perform well in terms of recommendation and are not suitable for subsequent large-scale recommendations. Therefore, it is necessary to set a smaller trial traffic to avoid a significant waste of recommendation resources. This application embodiment solves the aforementioned problems faced by medium-length video applications by setting personalized trial traffic. Therefore, medium-length video applications are suitable for using the video cold start method of this application embodiment. In medium-length video applications, the aforementioned target time is the time when the medium-length video application acquires the target video (the new medium-length video).

[0053] Step S202. Input the target video into the cold start behavior prediction model to obtain the cold start behavior prediction value. The cold start behavior prediction value indicates the probability that the target video will trigger the target behavior after being cold-started. The training samples of the cold start behavior prediction model are all constructed based on the sample videos released within the target time period and the target behaviors triggered by the sample videos within the target time period. The length of the target time period is a preset length, and the difference between the end time of the target time period and the target time is less than the preset length.

[0054] This application does not limit the preset length; the preset length can be set to a relatively short length, such as 10 days or 30 days. The target time period in this application is the time period closest to the target time, and each time period has the preset length.

[0055] In one exemplary implementation, the target behavior includes at least one individual behavior, and the cold start behavior prediction value includes at least one individual behavior prediction value. For example, the target behavior includes at least one of the following individual behaviors: exposure, click, and interaction. That is, the cold start behavior prediction model can comprehensively predict the probability of various individual behaviors being triggered by the target video, thereby comprehensively quantifying the consumption potential of the target video. Specifically, the cold start behavior prediction value may include exposure rate prediction value, click-through rate prediction value, and interaction rate prediction value.

[0056] This cold start behavior prediction model is trained using samples constructed from newly published videos within a short timeframe (target timeframe) adjacent to the target video (new video), thus distinguishing it from recommendation models in related technologies. The cold start behavior prediction model uses samples from a short timeframe, therefore the sample size is not too large, avoiding prediction inaccuracies caused by too many consumer-side samples. Furthermore, the target timeframe corresponding to the samples is highly close to the target time for acquiring the new video, avoiding prediction inaccuracies caused by long time intervals. This allows the cold start behavior prediction model to reflect the latest user interest and behavioral tendencies towards new videos.

[0057] This cold start behavior prediction model can be used in cold start scenarios. If this cold start behavior prediction model is not used in combination with video recommendation in non-cold start scenarios, nor is it used for recommending videos other than new videos, the purity of the cold start behavior prediction model can be guaranteed, and performance degradation caused by mixing can be avoided.

[0058] Step S203. Determine the cold start probe traffic based on the above cold start behavior prediction value; perform a cold start on the above target video based on the above cold start probe traffic.

[0059] In this embodiment, the cold start process for the target video based on the aforementioned cold start probe traffic can be the lower-level guarantee stage of the aforementioned tiered quantity guarantee scheme. Further, depending on the cold start effect, it can enter the support stage, which will not be elaborated upon in this embodiment. Unlike the lower-level guarantee stage of related technologies, the cold start probe traffic in this embodiment is determined based on the aforementioned cold start behavior prediction value, rather than being directly preset. Therefore, it is more reasonable, and the cold start effect of the lower-level guarantee stage is better.

[0060] This application does not limit the specific method for determining the cold start probe flow based on the above-mentioned cold start behavior prediction value. Generally speaking, the higher the cold start behavior prediction value, the larger the cold start probe flow. In an exemplary embodiment, please refer to... Figure 3 The diagram illustrates a flowchart of a cold start probe flow determination method according to an exemplary embodiment. The determination of the cold start probe flow based on the predicted cold start behavior includes:

[0061] Step S301. Determine the target trial flow.

[0062] The embodiments of this application do not limit the target trial flow, which has the same meaning as the preset trial flow in related technologies. For example, it can be preset to 200vv.

[0063] Step S302. Based on the above predicted cold start behavior value, update the above target trial flow to obtain the above cold start trial flow.

[0064] This application embodiment presets a target probe flow, corrects the target probe flow using cold start behavior prediction values, and uses the finally updated target probe flow as the cold start probe flow. Compared to related technologies that directly set the probe flow to 200vv for cold start based on experience, this application embodiment considers the indicative and corrective role of cold start behavior prediction values ​​on the cold start probe flow, thereby obtaining a more accurate cold start probe flow.

[0065] This application does not limit the update method. Generally speaking, if the cold start behavior prediction is high, the target probe traffic is updated to a larger value; conversely, if it is low, it is updated to a smaller value. In one embodiment, updating the target probe traffic based on the predicted cold start behavior value to obtain the cold start probe traffic includes:

[0066] If any of the predicted values ​​for the aforementioned individual behaviors is higher than the corresponding first threshold, a first weight is determined for the corresponding individual behavior; based on the product of the first weight and the target trial traffic, the target trial traffic is updated, and the first weight is greater than 1. The first threshold and the first weight in this embodiment can be set according to actual conditions and do not constitute an obstacle to implementation, so they will not be elaborated upon.

[0067] If any of the predicted values ​​for the aforementioned individual behaviors is lower than the corresponding second threshold, a second weight is determined for the corresponding individual behavior; based on the product of the second weight and the target trial flow, the target trial flow is updated, and the second weight is less than 1. The second threshold and the second weight in this embodiment can be set according to actual circumstances and do not constitute an obstacle to implementation, so they will not be elaborated upon further.

[0068] For example, the aforementioned single-behavioral prediction values ​​include exposure rate prediction value Mm, click-through rate prediction value Nn, and interaction rate prediction value Kk. The target trial traffic A is 200 vv. A is adjusted based on the exposure rate prediction value Mm, click-through rate prediction value Nn, and interaction rate prediction value Kk. For instance, if any one of Mm / Nn / Kk exceeds the corresponding first threshold, A = A * 1.2, where 1.2 is the first weight set based on experience and actual usage scenarios. If any one of Mm / Nn / Kk is below the corresponding second threshold, A = A * 0.8, where 0.8 is the second weight set based on experience and actual usage scenarios. By obtaining an updated and reasonable A, and thus obtaining cold start trial traffic, the video recommendation service can ensure sufficient distribution learning for newly published videos, which helps alleviate the Matthew effect and increases authors' enthusiasm for publishing new content.

[0069] To train the cold start behavior prediction model, embodiments of this application can construct training samples based on newly published videos and related feedback performance within a target time period. The training samples include at least one of the following types of samples:

[0070] The first sample set is constructed based on the first sample video and its exposure during the target time period. The second sample set is constructed based on the second sample video and its click data during the target time period. The third sample set is constructed based on the third sample video and its interaction data during the target time period. All three sample videos (first, second, and third) are newly published videos within the target time period. This application embodiment can obtain relevant sample sets from the perspective of target behaviors such as exposure, clicks, and interactions, enabling the cold start behavior prediction model to predict the probability of triggering exposure, clicks, and interactions for target videos.

[0071] In one embodiment, the first sample set is constructed by the following method: count the first videos with the number of exposures exceeding the first number threshold and the second videos with the number of exposures lower than the second number threshold within the target period, where the first videos and the second videos belong to the first sample videos; construct positive samples with the first videos and negative samples with the second videos to obtain the first sample set. The embodiments of the present application do not limit the magnitudes of the first number threshold and the second number threshold, which do not constitute an implementation obstacle.

[0072] Further, the second sample set is constructed by the following method: count the third videos with the click-through rate exceeding the first probability threshold and the fourth videos with the click-through rate lower than the second probability threshold within the target period, where the third videos and the fourth videos belong to the second sample videos; construct positive samples with the third videos and negative samples with the fourth videos to obtain the second sample set. The embodiments of the present application do not limit the magnitudes of the first probability threshold and the second probability threshold, which do not constitute an implementation obstacle.

[0073] Further, the third sample set is constructed by the following method: count the fifth videos with the number of interactions exceeding the third number threshold and the sixth videos with the number of interactions lower than the fourth number threshold within the target period, where the fifth videos and the sixth videos belong to the third sample videos; construct positive samples with the fifth videos and negative samples with the sixth videos to obtain the third sample set. The embodiments of the present application do not limit the magnitudes of the third number threshold and the fourth number threshold, which do not constitute an implementation obstacle.

[0074] For example, the embodiments of the present application can select the videos newly published in the recent N days (e.g., N = 30 days) and construct training samples based on their consumption situations in the video recommendation service, where the consumption situations specifically include the number of exposures, click-through rate, and number of interactions (including interactions such as likes and shares) of the videos. And determine M1, N, M2, N2, M3, N3 according to experience and specific scenarios. For the first sample set K1, the first videos with the number of exposures > M1 are used to construct positive samples, and the second videos with the number of exposures < N1 are used to construct negative samples. Usually, N1 < M1. For example, M1 = 1000 and N1 = 100. For the second sample set K2, the third videos with the click-through rate > M2 are used to construct positive samples, and the fourth videos with the click-through rate < N2 are used to construct negative samples. Usually, N2 < M2. For example, M2 = 0.05 and N2 = 0.01. For the third sample set K3, the fifth videos with the number of interactions > M3 are used to construct positive samples, and the sixth videos with the number of interactions < N3 are used to construct negative samples. Usually, N3 < M3. For example, M3 = 100 and N3 = 10.

[0075] In an exemplary embodiment, please refer to Figure 4It illustrates a schematic diagram of a cold start behavior prediction model structure in an exemplary implementation. Figure 1 The cold start behavior prediction model consists of three independent prediction models: exposure rate prediction model, click-through rate prediction model, and interaction rate prediction model. Accordingly, the training method for the above-mentioned cold start behavior prediction model includes at least one of the following:

[0076] The exposure rate prediction model is trained based on the first sample set mentioned above. The exposure rate prediction model belongs to the cold start behavior prediction model mentioned above.

[0077] The click-through rate prediction model is trained based on the second sample set mentioned above. The click-through rate prediction model belongs to the cold start behavior prediction model mentioned above.

[0078] The interaction rate prediction model is trained based on the third sample set mentioned above, and the interaction rate prediction model belongs to the cold start behavior prediction model mentioned above.

[0079] Based on the first sample set K1, the second sample set K2, and the third sample set K3, the exposure rate prediction model, click-through rate prediction model, and interaction rate prediction model can be trained independently, respectively. These models serve as independent components of the cold start behavior prediction model, enabling it to output predicted exposure rate, click-through rate, and interaction rate values. This application's embodiments do not elaborate on the training processes of each of the exposure rate prediction model, click-through rate prediction model, and interaction rate prediction model, and this does not constitute an obstacle to implementation.

[0080] In terms of specific details, the exposure rate prediction model, click-through rate prediction model, and interaction rate prediction model can all be built using deep neural networks. Parameters are adjusted using positive and negative samples based on binary classification cross-loss until training is complete. In the application phase, each model independently outputs its corresponding prediction results. This allows new videos with high exposure, high click-through rate, or high interaction potential to receive significant initial traffic, accumulating sufficient consumer performance data for subsequent amplified recommendations by the recommendation model.

[0081] In one exemplary implementation, please refer to Figure 5 It illustrates a schematic diagram of a cold start behavior prediction model structure in an exemplary implementation. Figure 2The aforementioned cold start behavior prediction model includes a feature extraction layer, a feature fusion layer, an exposure rate prediction layer, a click-through rate prediction layer, and an interaction rate prediction layer. The feature extraction layer is connected to the feature fusion layer, and the feature fusion layer is connected to the exposure rate prediction layer, the click-through rate prediction layer, and the interaction rate prediction layer, respectively. This cold start behavior prediction model is a comprehensive model, in which the exposure rate prediction layer, the click-through rate prediction layer, and the interaction rate prediction layer all reuse the feature extraction layer and the feature fusion layer, thereby saving storage and runtime space. The exposure rate prediction layer, the click-through rate prediction layer, and the interaction rate prediction layer each independently output the predicted exposure rate value, the predicted click-through rate value, and the predicted interaction rate value. This application does not limit the network structure of the feature extraction layer, the feature fusion layer, the exposure rate prediction layer, the click-through rate prediction layer, and the interaction rate prediction layer; convolutional networks or deep neural networks can be used, which does not constitute an obstacle to implementation.

[0082] In this embodiment of the application, for training... Figure 5 The cold start behavior prediction model uses a first sample set, a second sample set, and a third sample set as training samples. Positive samples in each of these sets include the corresponding video and a first preset value, which can be 1. Negative samples in each set include the corresponding video and a second preset value, which can be 0. For example, a positive sample in the first sample set consists of the first video and a value of 1, and a positive sample in the second sample set consists of the second video and a value of 0.

[0083] Please refer to Figure 6 The diagram illustrates the training process of a cold start behavior prediction model according to an exemplary embodiment. The cold start behavior prediction model is trained using the following method:

[0084] (1) Input the first sample video into the feature extraction layer to obtain the first feature output by the feature fusion layer; input the first feature into the exposure rate prediction layer to obtain the exposure rate prediction value; adjust the parameters of the exposure rate prediction layer based on the difference between the exposure rate prediction value and the preset value corresponding to the first sample video.

[0085] The aforementioned feature extraction layer is used to extract at least one of the following features: multimodal information extraction, static feature extraction, title extraction, and cover character extraction; the aforementioned feature fusion layer is used to fuse at least one of the following features: multimodal features, static features, title features, and cover character features. Therefore, the first feature can be the result of feature fusion of multiple features. In this embodiment, a multimodal model can be used to extract multimodal information to obtain multimodal features; the feature extraction layer can extract information including video author information, such as video author identifier information, video author level information, and video author fan information, to obtain static features. In this embodiment, video title information can be extracted to obtain title features; optical character recognition can be performed on the video cover to obtain cover character features.

[0086] This application does not limit the specific structure of the feature fusion layer; it can be implemented using a multilayer MLP. In the field of neural networks, a multilayer perceptron (MLP) is a type of feedforward artificial neural network. A multilayer MLP consists of multiple layers, including an input layer, one or more hidden layers, and an output layer. A multilayer MLP can be used to fuse multiple features to obtain the aforementioned first feature.

[0087] This application does not limit the specific method for adjusting the parameters of the exposure rate prediction layer based on the difference between the predicted exposure rate value and the preset value corresponding to the first sample video. If the first sample video is a positive sample, the corresponding preset value is 1; otherwise, it is 0. The loss can be obtained by calculating the Sigma ID cross-entropy loss. Sigma ID cross-entropy loss is a binary cross-entropy loss, a commonly used loss function in machine learning and deep learning, especially in binary classification problems.

[0088] Then, based on this loss, gradient descent is used to adjust the parameters of the exposure prediction layer. This embodiment does not limit the conditions for stopping parameter adjustment; for example, parameter tuning can be stopped when the loss is less than a preset loss threshold or the number of tuning attempts exceeds a preset number. The preset loss threshold and the preset number of attempts can be set according to actual conditions, and will not be elaborated here.

[0089] (2) Input the second sample video into the feature extraction layer to obtain the second feature output by the feature fusion layer; input the second feature into the click-through rate prediction layer to obtain the click-through rate prediction value; adjust the parameters of the click-through rate prediction layer based on the difference between the click-through rate prediction value and the preset value corresponding to the second sample video.

[0090] The training process of the click-through rate prediction layer is based on the same inventive concept as described in (1) above, and will not be repeated here.

[0091] (3) Input the third sample video into the feature extraction layer to obtain the third feature output by the feature fusion layer; input the third feature into the interaction rate prediction layer to obtain the interaction rate prediction value; adjust the parameters of the interaction rate prediction layer based on the difference between the interaction rate prediction value and the preset value corresponding to the third sample video.

[0092] The training process of the interaction rate prediction layer is based on the same inventive concept as described in (1) above, and will not be repeated here. In the embodiments of this application, the trained cold start behavior prediction model can output the exposure rate prediction value Mm, the click rate prediction value Nn, and the interaction rate prediction value Kk. All three prediction values ​​are numbers between 0 and 1.

[0093] Please refer to Figure 7 The diagram illustrates the entire cold start process of an exemplary implementation. First, training samples are collected during the target time period. Feature extraction is performed on these samples, and a cold start behavior prediction model based on the feature extraction results is trained, resulting in the trained cold start behavior prediction model. After acquiring the target video at the target time, the target video is input into the cold start behavior prediction model. After feature extraction and inference, the cold start behavior prediction model obtains a predicted cold start behavior value. Based on the predicted cold start behavior value, a cold start probe traffic can be determined, and a cold start is performed based on this probe traffic.

[0094] This solution can predict the potential of newly published target videos to trigger exposure, click-through rates, and interactions, thereby enabling targeted cold start testing traffic. It helps video recommendation services learn the most information about potentially successful videos with minimal traffic usage, allowing them to quickly launch popular new videos. This significantly promotes the new content ecosystem and motivates authors to publish.

[0095] The cold start solution of this application can be applied to various video cold start scenarios. Please refer to [link / reference]. Figure 8 This illustration shows an application scenario diagram of the cold start scheme according to an embodiment of this application. The application scenario is a medium-to-long video cold start scenario. In this scenario, new medium-to-long videos with a length of 1 minute or even 5 minutes or more are recommended to the user through a point-and-click approach. Figure 8 The system displays the recommended effects for new medium-length videos. Users can see the video's cover image and playback duration, with a brief description below the cover image. Figure 8The blurred portion at the bottom left corner represents the publishing account. Clicking on it triggers playback of the medium-length video, while clicking the "Watch Later" button adds it to the playlist for quick playback later. This cold start feature with medium-length videos satisfies users' need for thought-provoking and rewarding content.

[0096] The following are embodiments of the apparatus of this application, which can be used to execute embodiments of the method of this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the method of this application.

[0097] Please refer to Figure 9 The diagram illustrates a block diagram of a video cold start device according to an exemplary embodiment. This device has the function of implementing the video cold start method described above; this function can be implemented in hardware or by hardware executing corresponding software. The device can be a computer device or can be installed within a computer device. The device may include:

[0098] The video acquisition module 901 is used to acquire the target video, which is acquired at the target time.

[0099] Cold start module 902 is used to perform the following operations:

[0100] The target video is input into the cold start behavior prediction model to obtain the cold start behavior prediction value, which indicates the probability that the target video will trigger the target behavior after being cold-started.

[0101] Based on the above predicted cold start behavior values, determine the cold start trial flow rate;

[0102] Based on the cold start test traffic mentioned above, a cold start is performed on the aforementioned target video;

[0103] The training samples of the aforementioned cold start behavior prediction model are all constructed based on sample videos released within the target time period and the target behaviors triggered by the sample videos within the target time period. The length of the target time period is a preset length, and the difference between the end time of the target time period and the target time is less than the preset length.

[0104] In one exemplary embodiment, the cold start module 902 is configured to perform the following operations:

[0105] Determine the target and test the flow rate;

[0106] Based on the above predicted cold start behavior values, the above target trial flow is updated to obtain the above cold start trial flow.

[0107] In one exemplary embodiment, the target behavior includes at least one individual behavior, the cold start behavior prediction value includes at least one individual behavior prediction value, and the cold start module 902 is configured to perform the following operations:

[0108] If any of the above-mentioned individual behaviors is predicted to be higher than the corresponding first threshold, the first weight corresponding to the individual behavior is determined; based on the product of the first weight and the target trial traffic, the target trial traffic is updated, and the first weight is greater than 1;

[0109] If any of the above-mentioned individual behaviors is predicted to be lower than the corresponding second threshold, the second weight corresponding to the individual behavior is determined; based on the product of the above-mentioned second weight and the above-mentioned target trial traffic, the above-mentioned target trial traffic is updated, and the above-mentioned second weight is less than 1.

[0110] In one exemplary implementation, the target behavior includes at least one of the following individual behaviors: exposure, click, and interaction; the training samples include at least one of the following types of samples:

[0111] The first sample set is constructed based on the first sample video and the exposure of the first sample video during the target time period.

[0112] The second sample set is constructed based on the second sample video and the click data of the second sample video during the target time period.

[0113] The third sample set was constructed based on the third sample video and the interaction of the third sample video during the target time period.

[0114] In an exemplary embodiment, the first sample set is constructed by the following method: counting the number of exposures of a first video that exceeds a first threshold during the target time period, and the number of exposures of a second video that is lower than a second threshold during the target time period, wherein the first video and the second video belong to the first sample video; constructing positive samples with the first video and constructing negative samples with the second video to obtain the first sample set;

[0115] The second sample set was constructed using the following method: The third video with a click-through rate exceeding the first probability threshold and the fourth video with a click-through rate below the second probability threshold within the target time period were counted. The third and fourth videos belong to the second sample videos. Positive samples were constructed using the third video, and negative samples were constructed using the fourth video to obtain the second sample set.

[0116] The third sample set was constructed using the following method: the fifth video whose number of interactions exceeded the third threshold during the target time period, and the sixth video whose number of interactions was lower than the fourth threshold during the target time period. The fifth video and the sixth video belong to the third sample video. Positive samples were constructed using the fifth video, and negative samples were constructed using the sixth video to obtain the third sample set.

[0117] In one exemplary embodiment, the cold start module 902 is configured to perform the following operations:

[0118] The exposure rate prediction model is trained based on the first sample set mentioned above. The exposure rate prediction model belongs to the cold start behavior prediction model mentioned above.

[0119] The click-through rate prediction model is trained based on the second sample set mentioned above. The click-through rate prediction model belongs to the cold start behavior prediction model mentioned above.

[0120] The interaction rate prediction model is trained based on the third sample set mentioned above, and the interaction rate prediction model belongs to the cold start behavior prediction model mentioned above.

[0121] In one exemplary embodiment, the cold start behavior prediction model includes a feature extraction layer, a feature fusion layer, an exposure rate prediction layer, a click-through rate prediction layer, and an interaction rate prediction layer.

[0122] The aforementioned feature extraction layer is connected to the aforementioned feature fusion layer, and the aforementioned feature fusion layer is connected to the aforementioned exposure rate prediction layer, the aforementioned click-through rate prediction layer, and the aforementioned interaction rate prediction layer, respectively.

[0123] In an exemplary embodiment, the positive sample includes a corresponding video and a first preset value, the negative sample includes a corresponding video and a second preset value, and the cold start module 902 is used to perform the following operations:

[0124] The first sample video is input into the feature extraction layer to obtain the first feature output by the feature fusion layer; the first feature is input into the exposure rate prediction layer to obtain the exposure rate prediction value; based on the difference between the exposure rate prediction value and the preset value corresponding to the first sample video, the parameters of the exposure rate prediction layer are adjusted.

[0125] The second sample video is input into the feature extraction layer to obtain the second feature output by the feature fusion layer; the second feature is input into the click-through rate prediction layer to obtain the click-through rate prediction value; based on the difference between the click-through rate prediction value and the preset value corresponding to the second sample video, the parameters of the click-through rate prediction layer are adjusted.

[0126] The third sample video is input into the feature extraction layer to obtain the third feature output by the feature fusion layer; the third feature is input into the interaction rate prediction layer to obtain the interaction rate prediction value; based on the difference between the interaction rate prediction value and the preset value corresponding to the third sample video, the parameters of the interaction rate prediction layer are adjusted.

[0127] In one exemplary embodiment, the feature extraction layer is used to perform at least one of the following feature extractions: multimodal information extraction, static feature extraction, title extraction, and cover character extraction;

[0128] The feature fusion layer described above is used to fuse at least one of the following features: multimodal features, static features, title features, and cover character features.

[0129] It should be noted that the apparatus provided in the above embodiments is only illustrated by the division of the above functional modules when implementing its functions. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the apparatus and method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.

[0130] Please refer to Figure 10 It illustrates a structural block diagram of a computer device according to an exemplary embodiment for performing the video cold start method described above. Specifically:

[0131] Computer device 1000 includes a central processing unit (CPU) 1001, a system memory 1004 including random access memory (RAM) 1002 and read-only memory (ROM) 1003, and a system bus 1005 connecting the system memory 1004 and the CPU 1001. Computer device 1000 also includes a basic input / output system (I / O system) 1006 that facilitates information transfer between various devices within the computer, and a mass storage device 1007 for storing the operating system 1013, application programs 1014, and other program modules 1015.

[0132] The basic input / output system 1006 includes a display 1008 for displaying information and an input device 1009 for user input, such as a mouse or keyboard. Both the display 1008 and the input device 1009 are connected to the central processing unit 1001 via an input / output controller 1010 connected to the system bus 1005. The basic input / output system 1006 may also include the input / output controller 1010 for receiving and processing input from multiple other devices such as a keyboard, mouse, or electronic stylus. Similarly, the input / output controller 1010 also provides output to a display screen, printer, or other types of output devices.

[0133] Mass storage device 1007 is connected to central processing unit 1001 via a mass storage controller (not shown) connected to system bus 1005. Mass storage device 1007 and its associated computer-readable media provide non-volatile storage for computer device 1000. That is, mass storage device 1007 may include computer-readable media (not shown) such as hard disk or CD-ROM (Compact Disc Read-Only Memory) drive.

[0134] Without loss of generality, computer-readable media can include computer storage media and communication media. Computer storage media includes volatile and non-volatile, removable and non-removable media implemented using any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data. Computer storage media include RAM, ROM, EPROM (Erasable Programmable Read-Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), flash memory or other solid-state storage technologies, CD-ROM, DVD (Digital Video Disc) or other optical storage, magnetic tape cassettes, magnetic tape, disk storage, or other magnetic storage devices. Of course, those skilled in the art will recognize that computer storage media are not limited to the above-mentioned types. The system memory 1004 and mass storage device 1007 described above can be collectively referred to as memory.

[0135] According to various embodiments of this application, the computer device 1000 can also be connected to a remote computer on a network, such as the Internet. That is, the computer device 1000 can be connected to the network 1012 via the network interface unit 1011 connected to the system bus 1005, or the network interface unit 1011 can be used to connect to other types of networks or remote computer systems (not shown).

[0136] The aforementioned memory also includes a computer program stored in the memory and configured to be executed by one or more processors to implement the aforementioned video cold start method.

[0137] In an exemplary embodiment, a computer-readable storage medium is also provided, wherein at least one instruction, at least one program, code set, or instruction set is stored therein, wherein the at least one instruction, at least one program, code set, or instruction set is executed by a processor to implement the video cold start method described above.

[0138] Specifically, the video cold start method includes:

[0139] In one exemplary implementation, determining the cold start probe flow based on the predicted cold start behavior includes:

[0140] Determine the target and test the flow rate;

[0141] Based on the above predicted cold start behavior values, the above target trial flow is updated to obtain the above cold start trial flow.

[0142] In an exemplary implementation, the target behavior includes at least one individual behavior, the cold start behavior prediction value includes at least one individual behavior prediction value, and updating the target probe traffic based on the cold start behavior prediction value to obtain the cold start probe traffic includes:

[0143] If any of the above-mentioned individual behaviors is predicted to be higher than the corresponding first threshold, the first weight corresponding to the individual behavior is determined; based on the product of the first weight and the target trial traffic, the target trial traffic is updated, and the first weight is greater than 1;

[0144] If any of the above-mentioned individual behaviors is predicted to be lower than the corresponding second threshold, the second weight corresponding to the individual behavior is determined; based on the product of the above-mentioned second weight and the above-mentioned target trial traffic, the above-mentioned target trial traffic is updated, and the above-mentioned second weight is less than 1.

[0145] In one exemplary implementation, the target behavior includes at least one of the following individual behaviors: exposure, click, and interaction; the training samples include at least one of the following types of samples:

[0146] The first sample set is constructed based on the first sample video and the exposure of the first sample video during the target time period.

[0147] The second sample set is constructed based on the second sample video and the click data of the second sample video during the target time period.

[0148] The third sample set was constructed based on the third sample video and the interaction of the third sample video during the target time period.

[0149] In an exemplary embodiment, the first sample set is constructed by the following method: counting the number of exposures of a first video that exceeds a first threshold during the target time period, and the number of exposures of a second video that is lower than a second threshold during the target time period, wherein the first video and the second video belong to the first sample video; constructing positive samples with the first video and constructing negative samples with the second video to obtain the first sample set;

[0150] The second sample set was constructed using the following method: The third video with a click-through rate exceeding the first probability threshold and the fourth video with a click-through rate below the second probability threshold within the target time period were counted. The third and fourth videos belong to the second sample videos. Positive samples were constructed using the third video, and negative samples were constructed using the fourth video to obtain the second sample set.

[0151] The third sample set was constructed using the following method: the fifth video whose number of interactions exceeded the third threshold during the target time period, and the sixth video whose number of interactions was lower than the fourth threshold during the target time period. The fifth video and the sixth video belong to the third sample video. Positive samples were constructed using the fifth video, and negative samples were constructed using the sixth video to obtain the third sample set.

[0152] In one exemplary embodiment, the training method for the above-mentioned cold start behavior prediction model includes at least one of the following:

[0153] The exposure rate prediction model is trained based on the first sample set mentioned above. The exposure rate prediction model belongs to the cold start behavior prediction model mentioned above.

[0154] The click-through rate prediction model is trained based on the second sample set mentioned above. The click-through rate prediction model belongs to the cold start behavior prediction model mentioned above.

[0155] The interaction rate prediction model is trained based on the third sample set mentioned above, and the interaction rate prediction model belongs to the cold start behavior prediction model mentioned above.

[0156] In one exemplary embodiment, the cold start behavior prediction model includes a feature extraction layer, a feature fusion layer, an exposure rate prediction layer, a click-through rate prediction layer, and an interaction rate prediction layer.

[0157] The aforementioned feature extraction layer is connected to the aforementioned feature fusion layer, and the aforementioned feature fusion layer is connected to the aforementioned exposure rate prediction layer, the aforementioned click-through rate prediction layer, and the aforementioned interaction rate prediction layer, respectively.

[0158] In an exemplary implementation, the positive samples include the corresponding video and a first preset value, the negative samples include the corresponding video and a second preset value, and the cold start behavior prediction model is trained using the following method:

[0159] The first sample video is input into the feature extraction layer to obtain the first feature output by the feature fusion layer; the first feature is input into the exposure rate prediction layer to obtain the exposure rate prediction value; based on the difference between the exposure rate prediction value and the preset value corresponding to the first sample video, the parameters of the exposure rate prediction layer are adjusted.

[0160] The second sample video is input into the feature extraction layer to obtain the second feature output by the feature fusion layer; the second feature is input into the click-through rate prediction layer to obtain the click-through rate prediction value; based on the difference between the click-through rate prediction value and the preset value corresponding to the second sample video, the parameters of the click-through rate prediction layer are adjusted.

[0161] The third sample video is input into the feature extraction layer to obtain the third feature output by the feature fusion layer; the third feature is input into the interaction rate prediction layer to obtain the interaction rate prediction value; based on the difference between the interaction rate prediction value and the preset value corresponding to the third sample video, the parameters of the interaction rate prediction layer are adjusted.

[0162] In one exemplary embodiment, the feature extraction layer is used to perform at least one of the following feature extractions: multimodal information extraction, static feature extraction, title extraction, and cover character extraction;

[0163] The feature fusion layer described above is used to fuse at least one of the following features: multimodal features, static features, title features, and cover character features.

[0164] Optionally, the computer-readable storage medium may include: ROM (Read-Only Memory), RAM (Random Access Memory), SSD (Solid State Drive), or optical disc, etc. The random access memory may include ReRAM (Resistant Random Access Memory) and DRAM (Dynamic Random Access Memory).

[0165] In an exemplary embodiment, a computer program product or computer program is also provided, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the aforementioned video cold start method.

[0166] It should be understood that "multiple" as used herein refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. Furthermore, the step numbers described herein are merely illustrative of one possible execution order. In some other embodiments, the steps may not be executed in numerical order, such as two steps with different numbers being executed simultaneously, or two steps with different numbers being executed in the reverse order of the illustration. This application does not limit this.

[0167] In addition, in the specific embodiments of this application, data such as user information are involved. When the above embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0168] The above are merely exemplary embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application shall be included within the protection scope of this application.

Claims

1. A video cold start method, characterized in that, The method includes: Acquire the target video, which is acquired at a target time; The target video is input into the cold start behavior prediction model to obtain the cold start behavior prediction value, which indicates the probability that the target video will trigger the target behavior after being cold-started. Based on the predicted cold start behavior, determine the cold start trial flow; Based on the cold start probe traffic, a cold start is performed on the target video; The training samples of the cold start behavior prediction model are all constructed based on sample videos released within the target time period and the target behaviors triggered by the sample videos during the target time period. The length of the target time period is a preset length, and the difference between the end time of the target time period and the target time is less than the preset length.

2. The method according to claim 1, characterized in that, The step of determining the cold start probe flow based on the predicted cold start behavior includes: Determine the target and test the flow rate; Based on the predicted cold start behavior value, the target trial flow is updated to obtain the cold start trial flow.

3. The method according to claim 1 or 2, characterized in that, The target behavior includes at least one individual behavior, the cold start behavior prediction value includes at least one individual behavior prediction value, and updating the target probing traffic based on the cold start behavior prediction value to obtain the cold start probing traffic includes: If any of the predicted values ​​of a single behavior is higher than the corresponding first threshold, the first weight corresponding to the single behavior is determined; based on the product of the first weight and the target trial traffic, the target trial traffic is updated, and the first weight is greater than 1; If any of the predicted values ​​for a single behavior is lower than the corresponding second threshold, the second weight corresponding to the single behavior is determined; based on the product of the second weight and the target trial flow, the target trial flow is updated, and the second weight is less than 1.

4. The method according to claim 3, characterized in that, The target behavior includes at least one of the following individual behaviors: exposure, click, interaction; the training samples include at least one of the following types of samples: The first sample set is constructed based on the first sample video and the exposure of the first sample video during the target time period; The second sample set is constructed based on the second sample video and the click situation of the second sample video in the target time period; The third sample set is constructed based on the third sample video and the interaction of the third sample video during the target time period.

5. The method according to claim 4, characterized in that, The first sample set is constructed by: counting the number of exposures of the first video that exceeds the first threshold during the target time period, and the number of exposures of the second video that is lower than the second threshold during the target time period. The first video and the second video belong to the first sample video. Positive samples are constructed using the first video, and negative samples are constructed using the second video to obtain the first sample set; The second sample set is constructed by: identifying a third video whose click-through rate exceeds a first probability threshold and a fourth video whose click-through rate is below a second probability threshold within the target time period, wherein the third video and the fourth video belong to the second sample video; constructing positive samples with the third video and constructing negative samples with the fourth video to obtain the second sample set; The third sample set is constructed by: counting the fifth video whose number of interactions exceeds the third threshold during the target time period, and the sixth video whose number of interactions is lower than the fourth threshold during the target time period. The fifth video and the sixth video belong to the third sample video. Positive samples are constructed using the fifth video, and negative samples are constructed using the sixth video to obtain the third sample set.

6. The method according to claim 5, characterized in that, The training method for the cold start behavior prediction model includes at least one of the following: An exposure rate prediction model is trained based on the first sample set, and the exposure rate prediction model belongs to the cold start behavior prediction model; A click-through rate prediction model is trained based on the second sample set, and the click-through rate prediction model belongs to the cold start behavior prediction model. The interaction rate prediction model is trained based on the third sample set, and the interaction rate prediction model belongs to the cold start behavior prediction model.

7. The method according to claim 5, characterized in that, The cold start behavior prediction model includes a feature extraction layer, a feature fusion layer, an exposure rate prediction layer, a click-through rate prediction layer, and an interaction rate prediction layer. The feature extraction layer is connected to the feature fusion layer, and the feature fusion layer is connected to the exposure rate prediction layer, the click-through rate prediction layer, and the interaction rate prediction layer, respectively.

8. The method according to claim 7, characterized in that, The positive samples include the corresponding video and a first preset value, and the negative samples include the corresponding video and a second preset value. The cold start behavior prediction model is trained using the following method: The first sample video is input into the feature extraction layer to obtain the first feature output by the feature fusion layer; the first feature is input into the exposure rate prediction layer to obtain the exposure rate prediction value; the parameters of the exposure rate prediction layer are adjusted based on the difference between the exposure rate prediction value and the preset value corresponding to the first sample video. The second sample video is input into the feature extraction layer to obtain the second feature output by the feature fusion layer; The second feature is input into the click-through rate prediction layer to obtain the click-through rate prediction value; the parameters of the click-through rate prediction layer are adjusted based on the difference between the click-through rate prediction value and the preset value corresponding to the second sample video. The third sample video is input into the feature extraction layer to obtain the third feature output by the feature fusion layer; The third feature is input into the interaction rate prediction layer to obtain the interaction rate prediction value; Based on the difference between the predicted interaction rate and the preset value corresponding to the third sample video, the parameters of the interaction rate prediction layer are adjusted.

9. The method according to claim 8, characterized in that, The feature extraction layer is used to perform at least one of the following feature extractions: multimodal information extraction, static feature extraction, title extraction, and cover character extraction; The feature fusion layer is used to fuse at least one of the following features: multimodal features, static features, title features, and cover character features.

10. A video cold start device, characterized in that, The device includes: A video acquisition module is used to acquire a target video, which is acquired at a target time. The cold start module is used to perform the following operations: The target video is input into the cold start behavior prediction model to obtain the cold start behavior prediction value, which indicates the probability that the target video will trigger the target behavior after being cold-started. Based on the predicted cold start behavior, determine the cold start trial flow; Based on the cold start probe traffic, a cold start is performed on the target video; The training samples of the cold start behavior prediction model are all constructed based on sample videos released within the target time period and the target behaviors triggered by the sample videos during the target time period. The length of the target time period is a preset length, and the difference between the end time of the target time period and the target time is less than the preset length.

11. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing at least one instruction, at least one program, a code set, or an instruction set, the at least one instruction, the at least one program, the code set, or the instruction set being loaded and executed by the processor to implement the video cold start method as described in any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that, The storage medium stores at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by a processor to implement the video cold start method as described in any one of claims 1 to 9.

13. A computer program product, characterized in that, The computer program product includes computer instructions, the processor of the computer device reads the computer instructions, and the processor of the computer device executes the computer instructions to implement the video cold start method as described in any one of claims 1 to 9.

Images