Video popularity prediction method and device

Abstract

The present application provides a method and apparatus for predicting video popularity. The method comprises: extracting feature vectors corresponding to each of the multimodal features based on depth representation corresponding to the video data; inputting each of the feature vectors into a preset popularity prediction model based on a variational autoencoder (VAE); and using the corresponding popularity sequence output by the popularity prediction model as the popularity prediction result for the video data. The present application can effectively improve the utilization rate of video modalities, reduce information redundancy during the model training phase, and improve the accuracy of video popularity prediction results.

Description

Technical Field

[0001] The present application relates to the field of video information mining technology, and in particular to a method and device for predicting video popularity. Background Art

[0002] Video popularity prediction plays an important role in Internet data research. Predicting the popularity of videos can help video advertisers estimate the number of ad views in advance, assist video creators in clarifying the subsequent video content creation style and increase the creative incentives they can obtain, etc. It can also help scholars conduct research and analysis on Internet content dissemination and new media based on the results of video popularity prediction.

[0003] Currently, one of the existing methods for predicting video popularity usually uses video images or video title descriptions to predict the popularity of a video after it is released. However, due to the use of single-modal features, this method has a low utilization rate of the video modality and obviously cannot guarantee the accuracy of the video popularity prediction results. In order to solve this problem, the second existing method for predicting video popularity uses feature data containing additional video features to input support vector machines (SVM) or random forests (RF) to predict video popularity. However, due to the limitations brought by video creators and publishing platforms, video modalities have characteristics such as diversity or unstructured properties, which makes this method unable to effectively utilize these features. Therefore, the accuracy of the video popularity prediction results is still limited, and it is also easy to cause information redundancy in the model training stage. Summary of the Invention

[0004] In view of this, embodiments of the present application provide a video popularity prediction method and apparatus to eliminate or improve one or more defects in the prior art.

[0005] One aspect of the present application provides a video popularity prediction method, comprising:

[0006] Extracting feature vectors corresponding to the multimodal features based on depth representation corresponding to the video data;

[0007] Each of the feature vectors is input into a preset heat prediction model based on variational autoencoder VAE, and the heat sequence corresponding to the output of the heat prediction model is used as the heat prediction result of the video data.

[0008] In some embodiments of the present application, the heat prediction model includes: an input layer, a variational autoencoder VAE, and an output layer connected in sequence;

[0009] The hidden layer includes two fully connected layers connected in sequence;

[0010] The variational autoencoder VAE includes: a random DNN encoder, a hidden layer and a decoder based on a recurrent neural network;

[0011] The decoder includes a plurality of convolutional layers, a fully connected layer and an LSTM layer connected in sequence;

[0012] The input layer is used to receive a multimodal feature vector corresponding to the video data;

[0013] The random DNN encoder is used to obtain a probability representation corresponding to the multimodal feature vector output corresponding to the video data from the input layer;

[0014] The decoder is used to output a corresponding multidimensional feature vector according to the probability representation, and output a heat sequence having the same length as the dimension of the multidimensional vector through the output layer.

[0015] In some embodiments of the present application, each of the convolutional layers in the decoder includes: a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, and a fifth convolutional layer connected in sequence, and a maximum pooling layer is provided after the first convolutional layer, the second convolutional layer, and the fifth convolutional layer;

[0016] Among them, each convolutional layer and each fully connected layer in the heat prediction model is provided with a ReLu activation function, and the first fully connected layer and the LSTM layer in the heat prediction model are both provided with a random deactivation dropout layer.

[0017] In some embodiments of the present application, further comprising:

[0018] Get each historical video data;

[0019] Extracting multimodal feature vectors based on depth representation corresponding to each of the historical video data to obtain corresponding data sets;

[0020] Dividing the data set using a preset cross-validation method to obtain corresponding training sets and test sets;

[0021] Based on the training set, a machine learning model comprising an input layer, a variational autoencoder VAE and an output layer connected in sequence is trained, and based on the test results of the machine learning model for the test set, the machine learning model is adjusted to form a heat prediction model for predicting the heat sequence of video data.

[0022] In some embodiments of the present application, the multimodal features include: visual features, auditory features, textual features, and social features;

[0023] Wherein, the visual features include: global features, first emotional features and object features;

[0024] The auditory features include: Mel-frequency cepstral coefficients;

[0025] The text features include: semantic features and second emotional features;

[0026] The social features include: the number of fans and the number of followers of the video publisher of the video data, and an identifier indicating whether the video publisher is authenticated by the platform.

[0027] In some embodiments of the present application, extracting feature vectors corresponding to respective multimodal features based on depth representation corresponding to the video data includes:

[0028] Using a feature vector in a color histogram vector space as a global feature vector of the video data;

[0029] Extracting a first emotion feature vector of the video data based on a preset SentiBank data set;

[0030] Applying a preset densely connected convolutional network DenseNet to extract object feature vectors of the video data;

[0031] extracting Mel-frequency cepstral coefficients from the audio background in the video data;

[0032] Furthermore, text data corresponding to the video data, including a video title and a video description, is obtained, and the semantic feature vector and the second emotion feature vector are obtained based on the text data.

[0033] In some embodiments of the present application, obtaining the semantic feature vector and the second sentiment feature vector based on the text data includes:

[0034] Extracting corresponding semantic feature vectors from the text data based on pre-trained word2vec representation and bag-of-words model;

[0035] Based on the preset Chinese text processing tool SnowNLP, a sentiment score is assigned to the text data to obtain a corresponding second sentiment feature vector.

[0036] Another aspect of the present application provides a video popularity prediction device, comprising:

[0037] A multimodal feature extraction module is used to extract feature vectors corresponding to the multimodal features based on depth representation corresponding to the video data;

[0038] The heat prediction module is used to input each of the feature vectors into a preset heat prediction model based on the variational autoencoder VAE, and use the heat sequence corresponding to the output of the heat prediction model as the heat prediction result of the video data.

[0039] Another aspect of the present application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the video heat prediction method when executing the computer program.

[0040] Another aspect of the present application provides a computer-readable storage medium having a computer program stored thereon, which implements the video heat prediction method when executed by a processor.

[0041] The video heat prediction method provided by the present application extracts the feature vectors corresponding to the multimodal features based on depth representation corresponding to the video data; inputs each of the feature vectors into a preset heat prediction model based on variational autoencoder VAE, and uses the heat sequence corresponding to the output of the heat prediction model as the heat prediction result of the video data. By extracting the multimodal features corresponding to the video data and using the multimodal features to predict the heat of short videos, it can effectively improve the utilization rate of video modalities and improve the accuracy of video heat prediction results; at the same time, by using variational autoencoder VAE to predict the heat of short videos, and using the heat sequence as the prediction value to characterize the heat prediction of the video The measurement results can effectively eliminate the internal noise of multimodal features and the interference of external uncertainties, better represent the multimodality of videos, reduce information redundancy in the model training stage, and further improve the accuracy of video popularity prediction results. It can effectively help video advertisers estimate the number of advertisements in advance, assist video creators in clarifying the subsequent video content creation style and improve the creative incentives they can obtain, etc. It can also help scholars conduct research and analysis on Internet content dissemination and new media based on video popularity prediction results, and improve the user experience of video advertisers, video creators and researchers of Internet content dissemination and new media.

[0042] Additional advantages, purposes, and features of the present application will be described in part in the following description and will become apparent to those skilled in the art upon study of the following or may be learned from practice of the present application. The purposes and other advantages of the present application may be achieved and obtained by the structures specifically pointed out in the specification and drawings.

[0043] Those skilled in the art will understand that the purposes and advantages that can be achieved by the present application are not limited to the above specific description, and the above and other purposes that can be achieved by the present application will be more clearly understood based on the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The drawings described herein are intended to provide a further understanding of the present application, constitute a part of the present application, and do not constitute a limitation of the present application. The components in the drawings are not drawn to scale, but are only for the purpose of illustrating the principles of the present application. In order to facilitate the illustration and description of some parts of the present application, the corresponding parts in the drawings may be enlarged, that is, they may become larger than other components in the exemplary device actually manufactured according to the present application. In the drawings:

[0045] Figure 1 Schematic diagram of the overall process of the video popularity prediction method in one embodiment of the present application.

[0046] Figure 2 This is a schematic diagram of a specific flow chart of a video popularity prediction method in one embodiment of the present application.

[0047] Figure 3 This is a structural diagram of a video popularity prediction device in another embodiment of the present application.

[0048] Figure 4 This is a schematic diagram of the overall architecture of the short video popularity prediction model based on multimodal VAE in the application example of this application.

[0049] Figure 5 This is a detailed structural diagram of the variational autoencoder in the application example of this application. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail in conjunction with the embodiments and drawings. Here, the illustrative embodiments of this application and their descriptions are used to explain this application, but are not intended to limit this application.

[0051] It should also be noted here that in order to avoid obscuring the present application due to unnecessary details, the accompanying drawings only show structures and / or processing steps that are closely related to the scheme according to the present application, while other details that are not closely related to the present application are omitted.

[0052] It should be emphasized that the term "include / comprises" when used herein refers to the existence of features, elements, steps or components, but does not exclude the existence or addition of one or more other features, elements, steps or components.

[0053] It should also be noted that, unless otherwise specified, the term "connection" herein may refer not only to a direct connection but also to an indirect connection involving an intermediate.

[0054] Hereinafter, embodiments of the present application will be described with reference to the accompanying drawings. In the accompanying drawings, the same reference numerals represent the same or similar components, or the same or similar steps.

[0055] In one or more embodiments of the present application, the video popularity prediction may also be referred to as video playback volume prediction); and the video therein may refer to videos of various lengths, or may especially refer to short videos with a time limit. The short video may also be a type of user-generated content (UGC), wherein user-generated content (UGC) refers to users' original content that can be displayed or provided to other users through platforms such as the Internet. In addition to displaying original text, users can also publish original video content, which can be called short videos (or short videos, etc.), which are generally videos of less than 5 minutes in length that are circulated on new media on the Internet.

[0056] Taking short video popularity prediction as an example, most traditional video popularity prediction tasks have adopted a single-modal approach. For example, they predict a video's popularity after release based solely on the video's image or title description. However, this approach reduces the richness of video information and ignores the viewer's overall perception of the video. Existing single-modal research on social media popularity prediction using only visual and textual information has not yielded satisfactory results. Therefore, subsequent research on user-generated content (UGC) has refined the input of prediction models, moving from single text or video input to multimodal input. For example, existing techniques using single-modal feature extraction techniques have been the first to utilize visual modalities in the context of popularity prediction, addressing the problem of predicting the popularity of online images available on Flickr. This experiment, conducted on a dataset of over two million images, used image color histograms as visual features to feed a deep neural network and train a support vector machine (SVM) for popularity prediction. Predicting video popularity based on visual features allows video creators to predict the popularity of short videos before they are released, providing valuable insights for creators.

[0057] This method utilizes only a single video modality as input for video popularity prediction. The model relies on visual modalities as features and uses color histograms, a feature extraction method not based on a deep learning framework. Human perception of images relies solely on color perception, so this method has two drawbacks: 1) Feature extraction for images is limited to color histograms, ignoring semantic properties of the image itself, such as its content and emotional characteristics; 2) Using only the visual modality as input for the prediction model loses some video-related attributes. When video creators upload short videos to multimedia platforms, they also upload the video title, description, and other information.

[0058] To address the issues inherent in single-modal prediction, a feature-driven video popularity prediction method based on machine learning can also be employed. This method first extracts a large number of features related to UGC content, user profiles, or social networks, and trains machine learning models, such as support vector machines (SVMs) or random forests (RFs), to optimize the mapping function from the feature space to a predefined popularity space. This method primarily focuses on feature engineering techniques, and as long as the extracted features are valid, good performance can be achieved. On the other hand, to analyze the evolution of UGC popularity over time, generative methods utilize the temporal patterns of popularity curves in the early stages to fit autoregressive models with linear or nonlinear dynamics to predict future trends.

[0059] However, short videos can be comprehensively represented by leveraging a combination of visual, auditory, textual, and social features. Therefore, one of the key issues that arises in practical applications is the heterogeneous gap between features extracted from different views. Heterogeneous features extracted from different modalities show different aspects of a short video, and these modalities are interconnected in the hidden layer. In this case, due to the limitations imposed by short video producers and platforms, the additional information related to the short video, such as text descriptions, is diverse or unstructured, resulting in the unavailability of features extracted from specific views in many cases; traditional methods that fuse these features using simple concatenation or feature selection methods cannot well capture the semantic understanding of the features, and may therefore lead to information redundancy in the learning phase.

[0060] Therefore, a popularization prediction scheme based on a self-attention mechanism is also an option. A new method for predicting social media video popularity using a self-attention mechanism in video frames, text, and multimodal domains is proposed. This method demonstrates how to use visual and textual features to achieve more accurate predictions. This scheme proposes a method that combines Gradient Camera (GradCAM) with a soft-attention mechanism to visualize which parts of a video contribute to its popularity in the spatiotemporal domain. In addition to focusing on the visual impact of text in video descriptions, this scheme also demonstrates a complete system that improves the interpretability of social media video popularity prediction methods. However, while this popularization prediction method based on a self-attention mechanism proposes using visual and textual features as input for short video prediction, it ignores the impact of author influence on the dissemination of short videos. Furthermore, this method measures popularity using a single numerical value and fails to consider the effect of video play growth after a period of time, reducing the credibility of the popularization prediction.

[0061] In addition, a direct low-rank multimodal regression heat prediction scheme can also be considered, specifically a low-rank multimodal embedded framework to alleviate the problems of modal heterogeneity, interconnectedness and noise in short video heat prediction. By utilizing low-rank representation and multimodal learning, the framework effectively integrates all heterogeneous features extracted from different views into a common feature subspace and realizes enhanced robust feature representation for regression analysis. The framework also designs an effective optimization algorithm to solve the model. However, the direct low-rank multimodal regression heat prediction scheme has not broken the use of only single data as the heat representation of short videos for a long time, and compared with the variational autoencoder, the performance of the direct low-rank representation model is lower in the nMSE indicator.

[0062] Based on this, in response to the problems of information redundancy in the model training stage and low accuracy of video popularity prediction results in the above-mentioned existing technologies, the embodiments of the present application provide a video popularity prediction method based on multimodal feature extraction and variational autoencoder VAE, which can effectively improve the utilization rate of video modalities, reduce information redundancy in the model training stage, and thus improve the accuracy of video popularity prediction results, and improve the user experience of video advertising providers, video creators, and researchers of online content dissemination and new media.

[0063] Multimodal feature extraction, as a key input for prediction tasks, has long been a research hotspot for short video popularity prediction. Recent advances in deep learning and the availability of computing power have enabled multimedia data such as images and audio to be analyzed using the same deep learning framework as text, facilitating multimodal research. When selecting video modalities, visual, auditory, textual, and social information are the primary feature elements of interest. Based on existing popular technologies, the primary task of multimodal feature extraction is to extract features from these various feature elements using deep learning techniques and then combine them using specific methods. Traditional approaches to processing visual information utilize machine learning models, such as support vector machines (SVMs) and support vector virtual machines (SVRs). However, with technological advancements, deep learning has been widely applied to feature extraction tasks, such as using self-attention methods to semantically represent images and text.

[0064] In addition, popularity prediction technology based on variational autoencoders (VAEs) is one of the best performing techniques for popularity prediction. This technique primarily uses neural networks as encoders and decoders, and employs iterative optimization to learn the optimal encoding-decoding scheme. Therefore, in each iteration, video features are fed into an autoencoder structure (encoder followed by decoder). The encoded and decoded outputs are then compared with the initial data, and the network weights are updated via backpropagation of the error. Consequently, the entire autoencoder structure (encoder + decoder) creates a data bottleneck, ensuring that only the essential information can pass through the bottleneck and be reconstructed. A variational autoencoder can be defined as an autoencoder whose training is regularized to prevent overfitting and ensure that the latent space has good properties for data generation. The variational autoencoder for video popularity prediction consists of two parts: a multimodal encoder and a temporal decoder. In the encoder, multiple sets of vectors extracted from multimodal features are coupled using early feature fusion techniques to generate a distribution of feature vectors. In the decoder part, variational inference is used and KL divergence is introduced to predict the heat sequence of short videos based on the distribution of the input feature vector. In this way, the encoder and decoder parts are connected in series.

[0065] In one or more embodiments of the present application, Multi-modal refers to multimodality; VAE (Variational Auto-Encoder) refers to variational autoencoder; nMSE (normalized mean square error) refers to normalized mean square error; Spearman's Correlation refers to Spearman's correlation coefficient.

[0066] In this application, the main goal of multimodal feature extraction technology is to reduce the heterogeneity between modalities while maintaining the integrity of the independent semantics of each modality. This method has achieved excellent performance in deep learning.

[0067] The details are described in detail through the following examples.

[0068] Based on this, the embodiment of the present application provides a video popularity prediction method, see Figure 1 The video popularity prediction method that can be implemented by the NVMe controller specifically includes the following contents:

[0069] Step 100: Extract feature vectors corresponding to the multimodal features based on depth representation corresponding to the video data.

[0070] Step 200: Input each of the feature vectors into a preset heat prediction model based on a variational autoencoder (VAE), and use the heat sequence output by the heat prediction model as a heat prediction result of the video data.

[0071] In step 200, the heat sequence refers to the increment of the video playback volume within a period of time after the short video is released. The time interval can be adjusted according to the data set, such as the increment array of the video playback volume within seven days after the video is released, that is, a seven-dimensional array consisting of 7 numbers.

[0072] Specifically, in response to the problems of low prediction accuracy and insufficient modal application in the current video popularity prediction architecture in practical applications, steps 100 and 200 first proposed the use of a combination of multimodal feature extraction technology and variational autoencoders to achieve video popularity prediction. This not only solves the problem of low video modality utilization, but also introduces variational autoencoders into the model architecture of popularity prediction for the first time, creatively providing short videos with multiple sets of multi-dimensional, highly reliable feature vectors, and obtaining temporal information of video popularity, thereby improving the training accuracy of the model. The final experimental results show that this application has achieved strong competitive results on both public and self-built datasets, while providing a dataset that meets the video popularity prediction scenario and comparing its advantages with cutting-edge UGC content popularity prediction methods.

[0073] From the above description, it can be seen that the video popularity prediction method provided in the embodiment of the present application can effectively improve the utilization rate of video modality and improve the accuracy of video popularity prediction results by extracting multimodal features based on depth representation corresponding to video data and using multimodal features to predict the popularity of short videos; at the same time, by adopting variational autoencoder VAE to predict the popularity of short videos and using heat sequences as prediction values ​​to characterize the popularity prediction results of videos, it can effectively eliminate the internal noise of multimodal features and the interference of external uncertain factors, and can better represent the multimodality of videos, reduce information redundancy in the model training stage, and thus further improve the accuracy of video popularity prediction results. It can effectively help video advertisers estimate the number of advertisements in advance, assist video creators in clarifying the subsequent video content creation style and improving the creative incentives they can obtain, etc. It can also help scholars conduct research and analysis on Internet content dissemination and new media based on video popularity prediction results, and improve the user experience of video advertisers, video creators and researchers of online content dissemination and new media.

[0074] In order to further improve the application reliability and effectiveness of the popularity prediction model, in a video popularity prediction method provided in an embodiment of the present application, the popularity prediction model specifically includes the following contents:

[0075] The input layer, variational autoencoder VAE and output layer are connected in sequence;

[0076] The hidden layer includes two fully connected layers connected in sequence;

[0077] The variational autoencoder VAE includes: a random DNN encoder, a hidden layer and a decoder based on a recurrent neural network;

[0078] The decoder includes a plurality of convolutional layers, a fully connected layer and an LSTM layer connected in sequence;

[0079] The input layer is used to receive a multimodal feature vector corresponding to the video data;

[0080] The random DNN encoder is used to obtain a probability representation corresponding to the multimodal feature vector output corresponding to the video data from the input layer;

[0081] The decoder is used to output a corresponding multidimensional feature vector according to the probability representation, and output a heat sequence having the same length as the dimension of the multidimensional vector through the output layer.

[0082] Specifically, based on the statistical characteristics of short video heat sequences, this application uses a recurrent neural network as a decoder of the VAE structure. The working principle of the recurrent neural network RNN ​​is to save the output of a certain layer and feed it back to the input in order to predict the output of the decoder distribution. RNN can process time series data and accept current and previously received inputs. In addition, it is inspired by the long-term recurrent convolutional network (LRCN) architecture. It combines the sequential characteristics of the information displayed in the heat sequence, and the prediction effect is better than the method that only considers a single video heat value.

[0083] The recursive network architecture used in this application consists of two components: a convolutional neural network and a long short-term memory unit. After encoding, the application provides the integrated latent transformation as input to the decoder. The decoder network consists of seven layers and is a neural network with learnable parameters. Five convolutional layers, a fully connected layer, and an LSTM layer are connected after two fully connected layers of the hidden layer representation. The output of the network is a t-dimensional vector, the dimensionality of which depends on the length of the heat sequence.

[0084] From the above description, it can be seen that the video heat prediction method provided in the embodiment of the present application can effectively improve the application reliability and effectiveness of the heat prediction model, thereby better representing the multimodality of the video, reducing information redundancy in the model training stage, and further improving the accuracy of the video heat prediction results.

[0085] In order to further increase the translation invariance of the network and avoid overfitting, in a video heat prediction method provided in an embodiment of the present application, each of the convolutional layers in the decoder includes: a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, and a fifth convolutional layer connected in sequence, and a maximum pooling layer is provided after the first convolutional layer, the second convolutional layer, and the fifth convolutional layer;

[0086] Among them, each convolutional layer and each fully connected layer in the heat prediction model is provided with a ReLu activation function, and the first fully connected layer and the LSTM layer in the heat prediction model are both provided with a random deactivation dropout layer.

[0087] Specifically, to increase the network's translation invariance, we added max pooling layers after the first, second, and fifth convolutional layers. We used ReLu as the activation function and added it after each convolutional and fully connected layer. To avoid overfitting, we applied dropout layers after the first fully connected layer and the LSTM layer.

[0088] From the above description, it can be seen that the video heat prediction method provided in the embodiment of the present application can effectively increase the translation invariance of the network and avoid overfitting, thereby further improving the application reliability and effectiveness of the heat prediction model, so as to effectively improve the accuracy of the video heat prediction results.

[0089] In order to improve the training effectiveness and reliability of the popularity prediction model, in a video popularity prediction method provided in an embodiment of the present application, see Figure 2 The video popularity prediction method may further include the following steps before step 100:

[0090] Step 010: Obtain various historical video data.

[0091] Step 020: Extract the multimodal feature vectors based on depth representation corresponding to each of the historical video data to obtain the corresponding data set.

[0092] Step 030: Divide the data set using a preset cross-validation method to obtain corresponding training sets and test sets.

[0093] Step 040: Based on the training set, a machine learning model is trained, which includes an input layer, a variational autoencoder (VAE), and an output layer connected in sequence, and the machine learning model is adjusted based on the test result of the test set to form a heat prediction model for predicting the heat sequence of video data.

[0094] Specifically, for evaluation purposes, this application employed a 5-fold training-testing split (i.e., k-fold cross validation) to randomly partition the dataset. The 5-fold split refers to dividing the dataset into five parts (5 folds), with each part containing 20% ​​of the data. In each round of 5-fold cross validation, 9752 videos were used for training and 2438 for testing.

[0095] From the above description, it can be seen that the video heat prediction method provided in the embodiment of the present application can effectively improve the training effectiveness and reliability of the heat prediction model, and further improve the application reliability and effectiveness of the heat prediction model, so as to effectively improve the accuracy of the video heat prediction results.

[0096] In order to further improve the application reliability and effectiveness of multimodal features, in an embodiment of a video popularity prediction method provided in this application, the multimodal features include: visual features, auditory features, text features and social features;

[0097] Wherein, the visual features include: global features, first emotional features and object features;

[0098] The auditory features include: Mel-frequency cepstral coefficients;

[0099] The text features include: semantic features and second emotional features;

[0100] The social features include: the number of fans and the number of followers of the video publisher of the video data, and an identifier indicating whether the video publisher is authenticated by the platform.

[0101] Correspondingly, step 100 in the video popularity prediction method further specifically includes the following steps, which are executed in no particular order:

[0102] Step 110: using a feature vector in the color histogram vector space as a global feature vector of the video data;

[0103] Step 120: extracting a first emotion feature vector of the video data based on a preset SentiBank data set;

[0104] Step 130: Apply a preset densely connected convolutional network DenseNet to extract object feature vectors from the video data.

[0105] Specifically, global features: This application uses the color histogram as a global feature of the input video because it describes the surface properties of objects corresponding to the image or scene. For each input video, the first 5 seconds are selected as a representative sample, and 10 key frames are extracted from the sample on average. This application maps the colors in the key frames to a discrete color space containing 50 colors. Finally, this application obtains the feature vector in the color histogram 50-D vector space.

[0106] First, emotion features: This application extracts emotion features by training on the SentiBank dataset, a set of 2089 trained visual concept detectors that provide mid-level representations of emotion. This application uses the visual emotion ontology to train the SentiBank detector and collects all keyframes in the topic. Then, this application uses the SentiBank detector to obtain visual emotion features. Ultimately, this application obtains 2089-D visual emotion features for each video.

[0107] Object features: There is evidence that objects appearing in a video are one of the most important visual features for popularity prediction. This application uses a groundbreaking convolutional neural network model, the densely connected convolutional network (DenseNet), to perform video object detection. Specifically, this application constructs a multi-scale feature layer based on the single-shot multi-box detector (SSD) and uses the feature fusion module in the original SSD for target detection. The method of this application allows the features of all previous layers to be used as input to the next layer and realizes feature reuse. Therefore, this application generates a 1024-D vector for each video after average pooling of each key frame.

[0108] Step 140: Extract Mel-frequency cepstral coefficients from the audio background in the video data.

[0109] Specifically, acoustic information is inseparable from visual information in terms of conveying information and enhancing atmosphere, and these functions greatly affect the popularity of videos. Therefore, this application uses Mel-frequency cepstral coefficients (MFCC) extracted from the audio background as acoustic features to represent acoustic modalities. There are many ways to implement MFCC. The MFCC coefficients are a set of discrete cosine transforms (DCTs) calculated by transforming the output energy of the logarithmic compression filter. The discrete Fourier transform (DFT) signal is derived and processed by a perceptually spaced triangular filter bank. Unlike traditional MFCCs, this application uses DCT in a distributed mode to more closely track Markov features. Therefore, compared with traditional MFCCs, the count of feature vectors is reduced for the same number of filters. This application uses a window size of 12 and the differential and second-order differential of the MFCC features to generate 39-D acoustic features for each video.

[0110] Step 150: Obtain text data corresponding to the video data, including a video title and a video description, and obtain the semantic feature vector and the second emotion feature vector based on the text data.

[0111] Step 150 also specifically includes the following steps, which are executed in no particular order:

[0112] Step 151: extracting corresponding semantic feature vectors from the text data based on the pre-trained word2vec representation and bag-of-words model.

[0113] Specifically, semantic features: To capture the semantic representation of videos, this application develops deep features that represent text patterns. To obtain high-level semantic features, this application uses the skip-thought vectors of Kiros et al., using pre-trained word2vec representations and bag-of-words models, and uses an RNN to encode sentences into 4800-D vectors. Then, this application introduces fully connected layers, each followed by a tanh layer, to generate a 300-D text semantic feature vector.

[0114] Step 152: assigning a sentiment score to the text data based on a preset Chinese text processing tool SnowNLP to obtain a corresponding second sentiment feature vector.

[0115] Specifically, sentiment analysis of this large-scale text content can help better extract user sentiment towards events or topics. Text representation and convolutional neural network architectures are crucial components of sentiment analysis. Inspired by TextBlob, this application leverages the simplified Chinese text processing tool SnowNLP to assign a sentiment score to each video title and description—a decimal between 0 and 1 representing the degree of sentiment, ranging from negative to positive.

[0116] From the software level, this application also provides a video heat prediction device for executing all or part of the video heat prediction method, see Figure 3 The video heat prediction device specifically includes the following contents:

[0117] The multimodal feature extraction module 10 is used to extract feature vectors corresponding to the multimodal features based on depth representation corresponding to the video data.

[0118] The heat prediction module 20 is used to input each of the feature vectors into a preset heat prediction model based on variational autoencoder VAE, and use the heat sequence corresponding to the output of the heat prediction model as the heat prediction result of the video data.

[0119] The embodiment of the video heat prediction device provided in this application can be specifically used to execute the processing flow of the embodiment of the video heat prediction method in the above embodiment. Its functions will not be repeated here, and reference can be made to the detailed description of the embodiment of the above video heat prediction method.

[0120] The portion of the video popularity prediction device that performs video popularity prediction can be executed in a server, and in another practical application scenario, all operations can also be completed in the client device. The specific selection can be based on the processing capabilities of the client device and the limitations of the user's usage scenario. This application does not limit this. If all operations are completed in the client device, the client device may also include a processor for specific processing of video popularity prediction.

[0121] The client device may include a communication module (i.e., a communication unit) that can establish a communication connection with a remote server to implement data transmission with the server. The server may include a server on the task scheduling center side, and in other implementation scenarios, may also include a server on an intermediate platform, such as a server on a third-party server platform that has a communication link with the task scheduling center server. The server may include a single computer device, a server cluster consisting of multiple servers, or a server structure of a distributed device.

[0122] The server and the client device may communicate using any suitable network protocol, including network protocols that have not yet been developed as of the filing date of this application. Examples of such network protocols include TCP / IP, UDP / IP, HTTP, and HTTPS. Furthermore, examples of such network protocols include RPC (Remote Procedure Call Protocol) and REST (Representational State Transfer) protocols, which are used on top of the aforementioned protocols.

[0123] From the above description, it can be seen that the video heat prediction device provided in the embodiment of the present application can effectively improve the utilization rate of video modality and the accuracy of video heat prediction results by extracting multimodal features based on depth representation corresponding to video data and using multimodal features to predict the heat of short videos; at the same time, by adopting variational autoencoder VAE to predict the heat of short videos and using heat sequences as prediction values ​​to characterize the heat prediction results of videos, it can effectively eliminate the internal noise of multimodal features and the interference of external uncertain factors, and can better represent the multimodality of videos, reduce information redundancy in the model training stage, and thus further improve the accuracy of video heat prediction results. It can effectively help video advertisers estimate the number of advertisements in advance, assist video creators in clarifying the subsequent video content creation style and improving the creative incentives they can obtain, etc. It can also help scholars conduct research and analysis on Internet content dissemination and new media based on video heat prediction results, and improve the user experience of video advertisers, video creators and researchers of Internet content dissemination and new media.

[0124] To further illustrate this solution, taking short videos as an example, this application also provides a specific application example of a short video popularity prediction model based on multimodal feature extraction and variational autoencoder, see Figure 4 This application first characterizes different modalities through deep representation, then uses MLP as a random DNN encoder to derive the probabilistic representation of the intermediate layer, and finally uses the LSTM decoder to regress the temporal heat sequence. This model is not limited to the impact of temporal features on heat. The heat prediction method based on deep representation provides a better representation of multimodality, thereby more accurately predicting the future heat of UGC content. Compared with traditional heat definition methods, this method uses heat sequences as prediction values ​​to characterize the heat of UGC content, and uses VAE to eliminate internal noise and external uncertainty interference, thereby improving prediction accuracy.

[0125] The short video popularity prediction method based on the short video popularity prediction model specifically includes the following contents:

[0126] (1) Multimodal feature extraction

[0127] (1) Visual features

[0128] Global features: This application uses the color histogram as a global feature of the input video because it describes the surface properties of objects corresponding to the image or scene. For each input video, the first 5 seconds are selected as a representative sample, and 10 keyframes are extracted from this sample on average. The colors in the keyframes are mapped to a discrete color space containing 50 colors. Ultimately, this application obtains feature vectors in the color histogram 50-D vector space.

[0129] Emotional features: This application extracts emotional features by training on the SentiBank dataset, a set of 2089 trained visual concept detectors that provide mid-level representations of emotion. This application uses the visual emotion ontology to train the SentiBank detector and collects all keyframes in the topic. It then uses the SentiBank detector to obtain visual emotion features. Ultimately, this application obtains 2089-D visual emotion features for each video.

[0130] Object features: There is evidence that objects appearing in a video are one of the most important visual features for popularity prediction. This application uses a groundbreaking convolutional neural network model, the densely connected convolutional network (DenseNet), to perform video object detection. Specifically, this application constructs a multi-scale feature layer based on the single-shot multi-box detector (SSD) and uses the feature fusion module in the original SSD for target detection. The method of this application allows the features of all previous layers to be used as input to the next layer and realizes feature reuse. Therefore, this application generates a 1024-D vector for each video after average pooling of each key frame.

[0131] (2) Auditory characteristics

[0132] Acoustic information is inseparable from visual information in terms of conveying information and enhancing atmosphere, and these functions have greatly affected the popularity of videos. Therefore, this application uses Mel-frequency cepstral coefficients (MFCC) extracted from the audio background as acoustic features to represent acoustic modalities. There are many ways to implement MFCC. The MFCC coefficients are a set of discrete cosine transforms (DCTs) calculated by transforming the output energy of the logarithmic compression filter. The discrete Fourier transform (DFT) signal is derived and processed by a perceptually spaced triangular filter bank. Unlike traditional MFCCs, this application uses DCT in a distributed mode to more closely track Markov features. Therefore, compared with traditional MFCCs, the count of feature vectors is reduced for the same number of filters. This application uses a window size of 12 and the differential and second-order differential of the MFCC features to generate 39-D acoustic features for each video.

[0133] (3) Text features

[0134] Semantic Features: To capture the semantic representation of videos, we develop deep features that represent textual patterns. To obtain high-level semantic features, we use the skip-thought vectors from Kiros et al., using pre-trained word2vec representations and a bag-of-words model. We use an RNN to encode sentences into 4800-D vectors. We then introduce fully connected layers, each followed by a tanh layer, to generate a 300-D textual semantic feature vector.

[0135] Sentiment features: Sentiment analysis of such large-scale text content can help better extract user sentiment towards events or topics. Text representation and convolutional neural network architectures are crucial components of sentiment analysis. Inspired by TextBlob, this application leverages the simplified Chinese text processing tool SnowNLP to assign a sentiment score to each video title and description—a decimal between 0 and 1 representing the degree of sentiment, ranging from negative to positive.

[0136] (4) Social characteristics:

[0137] Number of followers: The number of followers of the video uploader.

[0138] Followers: The number of followers of the video uploader.

[0139] Platform certification: A binary number indicating whether the author is certified by the multimedia platform.

[0140] (2) Variational Autoencoder VAE

[0141] Based on the statistical characteristics of short video heat sequence, see Figure 5 , this application uses a recurrent neural network as a decoder of the VAE structure. The working principle of the recurrent neural network RNN ​​is to save the output of a certain layer and feed it back to the input in order to predict the output of the decoder distribution. RNN can process time series data and accept current and previously received inputs. In addition, it is inspired by the long-term recurrent convolutional network (LRCN) architecture. It combines the sequence characteristics of the information displayed in the heat sequence, and the prediction effect is better than the method that only considers a single video heat value. Among them, in Figure 5 In the equation, x represents multimodal video features; W represents random DNN encoder; σ represents the variance of the probability distribution of hidden features; μ represents the mean of the probability distribution of hidden features; W T Represents a decoder; Represents the predicted heat sequence.

[0142] The recursive network architecture used in this application consists of two components: a convolutional neural network and a long short-term memory unit. After encoding, the application provides the integrated latent transformation as input to the decoder. The decoder network consists of seven layers and is a neural network with learnable parameters. Five convolutional layers, a fully connected layer, and an LSTM layer are connected after two fully connected layers of the hidden layer representation. The output of the network is a t-dimensional vector, the dimensionality of which depends on the length of the heat sequence.

[0143] To increase the network's translation invariance, we added max pooling layers after the first, second, and fifth convolutional layers. We used ReLu as the activation function and added it after each convolutional and fully connected layer. To avoid overfitting, we applied dropout layers after the first fully connected layer and the LSTM layer.

[0144] (3) Model training

[0145] For evaluation purposes, this application randomly partitioned the dataset using a 5-fold training-testing split (i.e., k-fold cross validation). The 5-fold split refers to dividing the dataset into five parts (5 folds), with each part containing 20% ​​of the data. In each round of 5-fold cross validation, 9752 videos were used for training and 2438 for testing.

[0146] (IV) Clustering result analysis based on nMSE and Spearman coefficient and comparison with the prediction results of existing methods

[0147] The prediction results in Tables 1 and 2 show that our model outperforms existing popular methods. We draw the following conclusions from these results: First, models using multimodality achieve better performance than those using only a single modality. Because each modality not only shares information but also possesses modality-specific information, algorithms including CLSTM, MMVED, and MASSL outperform popular LRCN and attention models in terms of Spearman correlation coefficient. The unsatisfactory results of MvDA suggest that finding consistency across modalities is crucial. Since multimodal learning frameworks utilize latent spaces to better exploit the correlations between modalities, they naturally yield better latent space representations. A second observation is that the quality of the latent representation plays a crucial role in popularity prediction tasks. First, deep learning-based strategies outperform machine learning-based strategies. Compared to SVR, deep learning models achieve more promising results because their closer representations lead to less feature loss. On the other hand, autoencoder-based methods, including MMVED and MASSL, achieve more promising results than TMALL and CLSTM. In general, deep learning has shown greater power in multimodal practice.

[0148] Table 1

[0149] method nMSE SVR 1.132 MvD A 0.982 TMALL 0.979 MMVED 0.975 MASSL 0.968

[0150] Table 2

[0151] method Spearman SVR 0.384 Popularity LRCN 0.521 Attention models 0.607 MMVED 0.742 MASSL 0.781

[0152] The prediction results in Table 3 show that our model outperforms existing popular methods. We draw the following conclusions from these results: First, models using multimodality achieve better performance than those using only a single modality. Because each modality not only shares information but also possesses modality-specific information, algorithms including CLSTM, MMVED, and MASSL outperform popular LRCN and attention models in terms of Spearman correlation coefficient. The unsatisfactory results of MvDA suggest that finding consistency across modalities is crucial. Since multimodal learning frameworks utilize latent spaces to better exploit the correlations between modalities, they naturally yield better latent space representations. A second observation is that the quality of the latent representation plays a crucial role in popularity prediction tasks. First, deep learning-based strategies outperform machine learning-based strategies. Compared to SVR, deep learning models achieve more promising results because their closer representations lead to less feature loss. On the other hand, autoencoder-based methods, including MMVED and MASSL, achieve more promising results than TMALL and CLSTM. In general, deep learning has shown greater power in multimodal practice.

[0153] Table 3

[0154] Multimodal combination nMSE Spearman T+V+A 0.975 0.779 T+A+S 0.973 0.781 T+V+S 0.973 0.782 T+A+S 0.971 0.781 T+V+A+S 0.968 0.781

[0155] In summary, the application example of this application uses multimodal feature extraction technology and variational autoencoder network to extract a set of feature vectors from the multimodality of short videos and predict the popularity of short videos by encoding and decoding the time series through VAE. The architecture proposed in this application is not limited to the impact of time series features on popularity. The popularity prediction method based on deep representation better represents multimodality, thereby more accurately predicting the future popularity of UGC content. Compared with the traditional popularity definition method, this application uses the heat sequence as the prediction value to characterize the popularity of UGC content, and uses VAE to eliminate internal noise and external uncertainty factors, thereby improving the prediction accuracy.

[0156] An embodiment of the present application also provides an electronic device (i.e., an electronic device), which may include a processor, a memory, a receiver, and a transmitter, wherein the processor is used to execute the video heat prediction method mentioned in the above embodiment, wherein the processor and the memory may be connected via a bus or other means, taking a bus connection as an example. The receiver may be connected to the processor and the memory via a wired or wireless manner. The electronic device may receive real-time motion data from sensors in the wireless multimedia sensor network, and receive original video sequences from the video acquisition device.

[0157] The processor may be a central processing unit (CPU). The processor may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or a combination of the above chips.

[0158] The memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer executable programs, and modules, such as the program instructions / modules corresponding to the video popularity prediction method in the embodiments of the present application. The processor executes the non-transitory software programs, instructions, and modules stored in the memory to perform various functional applications and data processing of the processor, thereby implementing the video popularity prediction method in the above method embodiment.

[0159] The memory may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created by the processor, etc. In addition, the memory may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory may optionally include a memory remotely located relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0160] The one or more modules are stored in the memory, and when executed by the processor, perform the video heat prediction method in the embodiment.

[0161] In some embodiments of the present application, the user equipment may include a processor, a memory and a transceiver unit, and the transceiver unit may include a receiver and a transmitter. The processor, memory, receiver and transmitter may be connected through a bus system. The memory is used to store computer instructions, and the processor is used to execute the computer instructions stored in the memory to control the transceiver unit to send and receive signals.

[0162] As an implementation method, the functions of the receiver and transmitter in this application can be considered to be implemented through a transceiver circuit or a dedicated transceiver chip, and the processor can be considered to be implemented through a dedicated processing chip, a processing circuit or a general-purpose chip.

[0163] As another implementation method, it is possible to use a general-purpose computer to implement the server provided in the embodiments of the present application. That is, the program code for implementing the functions of the processor, receiver, and transmitter is stored in a memory, and the general-purpose processor implements the functions of the processor, receiver, and transmitter by executing the code in the memory.

[0164] The present application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the aforementioned video popularity prediction method. The computer-readable storage medium can be a tangible storage medium, such as a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a floppy disk, a hard disk, a removable storage disk, a CD-ROM, or any other form of storage medium known in the art.

[0165] It should be understood by those skilled in the art that the various exemplary components, systems and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software or a combination of the two. Whether it is specifically performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application. When implemented in hardware, it can be, for example, an electronic circuit, an application-specific integrated circuit (ASIC), appropriate firmware, a plug-in, a function card, etc. When implemented in software, the elements of this application are programs or code segments that are used to perform the required tasks. The program or code segment can be stored in a machine-readable medium, or transmitted on a transmission medium or a communication link via a data signal carried in a carrier.

[0166] It should be understood that the present application is not limited to the specific configurations and processes described above and illustrated in the figures. For the sake of brevity, a detailed description of known methods is omitted here. In the above embodiments, several specific steps are described and illustrated as examples. However, the method process of the present application is not limited to the specific steps described and illustrated. Those skilled in the art can make various changes, modifications, and additions, or change the order of the steps after understanding the spirit of the present application.

[0167] In this application, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or replace features of other embodiments.

[0168] The above description is merely a preferred embodiment of the present application and is not intended to limit the present application. Those skilled in the art will appreciate that various modifications and variations of the present embodiment are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present application shall be included within the scope of protection of the present application.

Claims

1. A video popularity prediction method, characterized in that: include: Extracting feature vectors corresponding to the multimodal features based on depth representation corresponding to the video data; Input each of the feature vectors into a preset heat prediction model based on a variational autoencoder (VAE), and use the heat sequence corresponding to the output of the heat prediction model as a heat prediction result of the video data; The heat prediction model includes: an input layer, a variational autoencoder (VAE), and an output layer connected in sequence; The variational autoencoder VAE includes: a random DNN encoder, a hidden layer and a decoder based on a recurrent neural network; the hidden layer includes two fully connected layers connected in sequence; The decoder includes a plurality of convolutional layers, a fully connected layer and an LSTM layer connected in sequence; The input layer is used to receive a multimodal feature vector corresponding to the video data; The random DNN encoder is used to obtain a probability representation corresponding to the multimodal feature vector output corresponding to the video data from the input layer; The decoder is used to output a corresponding multidimensional feature vector according to the probability representation, and output a heat sequence having the same length as the dimension of the multidimensional feature vector through the output layer; Each of the convolutional layers in the decoder includes: a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, and a fifth convolutional layer connected in sequence, and a maximum pooling layer is provided after each of the first convolutional layer, the second convolutional layer, and the fifth convolutional layer; Among them, each convolutional layer and each fully connected layer in the heat prediction model is provided with a ReLu activation function, and the first fully connected layer and the LSTM layer in the heat prediction model are both provided with a random deactivation dropout layer.

2. The video popularity prediction method according to claim 1, characterized in that: Also includes: Get each historical video data; Extracting multimodal feature vectors based on depth representation corresponding to each of the historical video data to obtain corresponding data sets; Dividing the data set using a preset cross-validation method to obtain corresponding training sets and test sets; Based on the training set, a machine learning model comprising an input layer, a variational autoencoder VAE and an output layer connected in sequence is trained, and based on the test results of the machine learning model for the test set, the machine learning model is adjusted to form a heat prediction model for predicting the heat sequence of video data.

3. The video popularity prediction method according to claim 1 or 2, characterized in that: The multimodal features include: visual features, auditory features, text features and social features; Wherein, the visual features include: global features, first emotional features and object features; The auditory features include: Mel-frequency cepstral coefficients; The text features include: semantic features and second emotional features; The social features include: the number of fans and the number of followers of the video publisher of the video data, and an identifier indicating whether the video publisher is authenticated by the platform.

4. The video popularity prediction method according to claim 3, characterized in that: The extracting of feature vectors corresponding to the multimodal features based on depth representation corresponding to the video data includes: Using a feature vector in a color histogram vector space as a global feature vector of the video data; Extracting a first emotion feature vector of the video data based on a preset SentiBank data set; Applying a preset densely connected convolutional network DenseNet to extract object feature vectors of the video data; extracting Mel-frequency cepstral coefficients from the audio background in the video data; Furthermore, text data corresponding to the video data, including a video title and a video description, is obtained, and a semantic feature vector and a second emotion feature vector are obtained based on the text data.

5. The video popularity prediction method according to claim 4, characterized in that: The obtaining of the semantic feature vector and the second emotional feature vector according to the text data includes: Extracting corresponding semantic feature vectors from the text data based on pre-trained word2vec representation and bag-of-words model; Based on the preset Chinese text processing tool SnowNLP, a sentiment score is assigned to the text data to obtain a corresponding second sentiment feature vector.

6. A video popularity prediction device, characterized in that: include: A multimodal feature extraction module is used to extract feature vectors corresponding to the multimodal features based on depth representation corresponding to the video data; A popularity prediction module is used to input each of the feature vectors into a preset popularity prediction model based on a variational autoencoder (VAE), and use the corresponding heat sequence output by the heat prediction model as a popularity prediction result of the video data; The heat prediction model includes: an input layer, a variational autoencoder (VAE), and an output layer connected in sequence; The variational autoencoder VAE includes: a random DNN encoder, a hidden layer and a decoder based on a recurrent neural network; the hidden layer includes two fully connected layers connected in sequence; The decoder includes a plurality of convolutional layers, a fully connected layer and an LSTM layer connected in sequence; The input layer is used to receive a multimodal feature vector corresponding to the video data; The random DNN encoder is used to obtain a probability representation corresponding to the multimodal feature vector output corresponding to the video data from the input layer; The decoder is used to output a corresponding multidimensional feature vector according to the probability representation, and output a heat sequence having the same length as the dimension of the multidimensional feature vector through the output layer; Each of the convolutional layers in the decoder includes: a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, and a fifth convolutional layer connected in sequence, and a maximum pooling layer is provided after each of the first convolutional layer, the second convolutional layer, and the fifth convolutional layer; Among them, each convolutional layer and each fully connected layer in the heat prediction model is provided with a ReLu activation function, and the first fully connected layer and the LSTM layer in the heat prediction model are both provided with a random deactivation dropout layer.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the video heat prediction method according to any one of claims 1 to 5 is implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the video heat prediction method according to any one of claims 1 to 5 is implemented.

Images