Video popularity assessment method, device and equipment and computer readable storage medium

Abstract

The invention discloses a video popularity assessment method, device and equipment and a computer readable storage medium, and the method comprises the steps: carrying out the feature extraction of the video content of a target video, and obtaining a video feature; the video features comprise visual features, audio features and text features; performing sentiment analysis on the user comment data of the target video to obtain user sentiment features; the user emotion features comprise user emotion distribution features and user comment features; performing dynamic trajectory analysis on the user interaction behavior of the target video to obtain user behavior characteristics; the user behavior characteristics comprise user behavior evolution characteristics; performing cross-modal fusion processing on the video features, the user emotion features and the user behavior features to obtain fusion features; and determining a popularity evaluation result of the target video according to the fusion feature. By applying the technical scheme provided by the invention, dynamic, accurate and interpretable evaluation of the video propagation potential can be realized.

Description

Technical Field

[0001] This application relates to the field of multimedia technology, and in particular to a video heat assessment method, apparatus, electronic device, and computer-readable storage medium. Background Technology

[0002] With the widespread adoption of mobile internet and smart devices, short video platforms and social media have become major channels for people to obtain information, entertainment, and social interaction. According to relevant statistics, the number of short video users worldwide has exceeded several billion, with hundreds of millions of videos generated daily. In this massive content environment, accurately assessing video popularity to provide a reliable basis for content distribution, recommendation systems, and business decisions has become a core challenge for platform operations.

[0003] Video popularity assessment refers to measuring the dissemination effectiveness and popularity of video content among a user group using quantitative indicators. Currently, related technologies mainly rely on explicit behavioral data such as play counts, likes, share rates, and completion times, using weighted calculations or machine learning regression to predict the video's dissemination potential. While these methods offer advantages such as quantifiable data and standardized structures, they generally neglect the implicit emotional feedback generated by users during viewing, such as unstructured emotional signals like the semantic bias of comments, the emotional density of bullet comments, and the distribution of emoticons. This leads to a disconnect between the assessment results and the users' actual psychological experience, making it difficult to accurately reflect the video's actual dissemination effectiveness.

[0004] Therefore, how to achieve more accurate and effective video popularity assessment is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] The purpose of this application is to provide a video popularity assessment method that can achieve a dynamic, accurate, and interpretable assessment of the video's dissemination potential; another purpose of this application is to provide a video popularity assessment device, electronic device, and computer-readable storage medium, all of which have the aforementioned beneficial effects.

[0006] Firstly, this application discloses a method for evaluating video popularity, including:

[0007] Feature extraction is performed on the video content of the target video to obtain video features; the video features include visual features, audio features, and text features;

[0008] Sentiment analysis is performed on the user comment data of the target video to obtain user sentiment features; the user sentiment features include user sentiment distribution features and user comment features.

[0009] Dynamic trajectory analysis is performed on the user interaction behavior of the target video to obtain user behavior characteristics; the user behavior characteristics include user behavior evolution characteristics;

[0010] The video features, user emotion features, and user behavior features are subjected to cross-modal fusion processing to obtain fused features;

[0011] The heat assessment result of the target video is determined based on the fusion features.

[0012] Optionally, feature extraction is performed on the video content of the target video to obtain video features, including:

[0013] Based on the video content of the target video, video frames are determined, and the video frames are processed using a first network model to obtain the visual features; the visual features include camera movement patterns, scene switching frequency, density of main objects, visual rhythm, and dynamic changes;

[0014] The audio signal is determined based on the video content of the target video, and the Mel frequency cepstral coefficients are extracted from the audio signal. The short-time energy and zero-crossing rate of each audio frame in the audio signal are calculated, and the Mel frequency cepstral coefficients, each short-time energy, and the zero-crossing rate are used as the audio features.

[0015] Text data is determined based on the video content of the target video, and the inverse document frequency weight of each semantic unit in the text data is calculated. Topic modeling is performed based on each inverse document frequency weight to obtain the topic distribution probability as the text feature.

[0016] Optionally, sentiment analysis is performed on the user comment data of the target video to obtain user sentiment characteristics, including:

[0017] The second network model is used to process the user comment data of the target video to obtain the sentiment type of each user comment data, and the user sentiment distribution characteristics are determined based on the sentiment type of each user comment data; the sentiment type includes positive sentiment, negative sentiment, and neutral sentiment;

[0018] The total number of user comments and the comment density growth rate are determined based on the user comment data of the target video, and the total number of user comments and the comment density growth rate are used as the user comment features.

[0019] Optionally, dynamic trajectory analysis is performed on the user interaction behavior of the target video to obtain user behavior characteristics, including:

[0020] The user interaction behavior of the target video is sampled according to a preset time window to obtain the time sequence data of each user interaction behavior.

[0021] The user behavior evolution characteristics are obtained by modeling the temporal dependencies of the time-series data of each user interaction behavior using a third network model.

[0022] Optionally, the video features, the user sentiment features, and the user behavior features are subjected to cross-modal fusion processing to obtain fused features, including:

[0023] A multi-head self-attention network is used to calculate the weights of the video features, the user emotion features, and the user behavior features to obtain the attention weights of the video features, the user emotion features, and the user behavior features.

[0024] The video features, user emotion features, and user behavior features are weighted and fused using the attention weights to obtain the fused features.

[0025] Optionally, determining the heat assessment result of the target video based on the fusion features includes:

[0026] The fused features are processed using a regressor network to obtain the heat evaluation value of the target video;

[0027] The target popularity level to which the popularity assessment value belongs is determined according to the popularity level classification rules, and the target popularity level is used as the popularity assessment result of the target video.

[0028] Optionally, the video popularity assessment method further includes:

[0029] Calculate the time derivative of the heat assessment value;

[0030] When the absolute value of the time derivative exceeds a preset threshold, an abnormal heat alarm is triggered.

[0031] The rate of change in the sentiment distribution of user comments is determined based on the user comment data.

[0032] The fluctuation pattern of user behavior indicators is determined based on the user interaction behavior.

[0033] The abnormal pattern corresponding to the abnormal popularity alarm is determined and output based on the change rate of the user comment sentiment distribution and the fluctuation pattern of the user behavior indicators.

[0034] Secondly, this application discloses a video heat assessment device, comprising:

[0035] The first processing module is used to extract features from the video content of the target video to obtain video features; the video features include visual features, audio features, and text features.

[0036] The second processing module is used to perform sentiment analysis on the user comment data of the target video to obtain user sentiment features; the user sentiment features include user sentiment distribution features and user comment features.

[0037] The third processing module is used to perform dynamic trajectory analysis on the user interaction behavior of the target video to obtain user behavior features; the user behavior features include user behavior evolution features.

[0038] The fusion module is used to perform cross-modal fusion processing on the video features, the user emotion features, and the user behavior features to obtain fused features;

[0039] A determination module is used to determine the heat assessment result of the target video based on the fusion features.

[0040] Thirdly, this application discloses an electronic device, including:

[0041] Memory, used to store computer programs;

[0042] A processor for executing the computer program to implement any of the video heat assessment methods described above.

[0043] Fourthly, this application discloses a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any of the video heat assessment methods described above.

[0044] This application provides a method for evaluating video popularity, comprising: extracting features from the video content of a target video to obtain video features; the video features include visual features, audio features, and text features; performing sentiment analysis on user comment data of the target video to obtain user sentiment features; the user sentiment features include user sentiment distribution features and user comment features; performing dynamic trajectory analysis on user interaction behavior of the target video to obtain user behavior features; the user behavior features include user behavior evolution features; performing cross-modal fusion processing on the video features, the user sentiment features, and the user behavior features to obtain fused features; and determining the popularity evaluation result of the target video based on the fused features.

[0045] By applying the technical solution provided in this application, for a target video, its video features, user emotional features, and user behavioral features can be extracted and fused together to achieve video popularity assessment. It can be seen that this technical solution takes into account the implicit emotional feedback generated by users during video viewing to achieve video popularity assessment, which helps to improve the accuracy of video popularity assessment results. It realizes a dynamic, accurate, and interpretable assessment of the video's dissemination potential, and can provide more reliable data basis for subsequent content distribution, recommendation systems, and business decisions.

[0046] The video heat assessment device, electronic device, and computer-readable storage medium provided in this application also have the above-mentioned technical effects, and will not be described in detail here. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the prior art and the embodiments of this application, the accompanying drawings used in the description of the prior art and the embodiments of this application will be briefly introduced below. Of course, the accompanying drawings described below with respect to the embodiments of this application are only a part of the embodiments in this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort, and such other drawings also fall within the protection scope of this application.

[0048] Figure 1 A flowchart illustrating a video popularity assessment method provided in an embodiment of this application;

[0049] Figure 2 This is a schematic diagram of the structure of a video heat assessment system provided in an embodiment of this application;

[0050] Figure 3 This is a schematic diagram of the structure of a video content feature extraction module provided in an embodiment of this application;

[0051] Figure 4 This is a schematic diagram of the structure of a user comment sentiment analysis module provided in an embodiment of this application;

[0052] Figure 5 This is a schematic diagram of the structure of a viewing behavior timing modeling module provided in an embodiment of this application;

[0053] Figure 6 This is a schematic diagram of the structure of a comprehensive heat decision module provided in an embodiment of this application;

[0054] Figure 7 This is a schematic diagram of the structure of a heat level classification module provided in an embodiment of this application;

[0055] Figure 8This is a schematic diagram of the structure of an online model update mechanism provided in an embodiment of this application;

[0056] Figure 9 This is a schematic diagram of the structure of an abnormal heat detection submodule provided in an embodiment of this application;

[0057] Figure 10 This is a schematic diagram of a batch parallel evaluation architecture provided in an embodiment of this application;

[0058] Figure 11 This is a schematic diagram of the structure of a visualization output module provided in an embodiment of this application;

[0059] Figure 12 This is a schematic diagram of the structure of a video heat assessment device provided in an embodiment of this application;

[0060] Figure 13 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0061] The core of this application is to provide a video popularity assessment method that can achieve a dynamic, accurate, and interpretable assessment of a video's dissemination potential. Another core aspect of this application is to provide a video popularity assessment device, electronic device, and computer-readable storage medium, all of which have the aforementioned beneficial effects.

[0062] To provide a clearer and more complete description of the technical solutions in the embodiments of this application, the technical solutions in the embodiments of this application will be described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0063] This application provides a method for evaluating video popularity.

[0064] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating a video popularity assessment method provided in an embodiment of this application. The video popularity assessment method may include, but is not limited to, the following S101~S105.

[0065] S101: Extract features from the video content of the target video to obtain video features; video features include visual features, audio features, and text features.

[0066] This step aims to extract video features from the target video. These features can be extracted based on the video content of the target video. The video features can include visual features, audio features, and text features. Visual features are the feature information extracted from the video frames of the target video, audio features are the feature information extracted from the audio signals of the target video, and text features are the feature information extracted from the text data of the target video.

[0067] In one embodiment of this application, feature extraction of the video content of the target video to obtain video features may include:

[0068] Based on the video content of the target video, the video frames are determined, and the video frames are processed using the first network model to obtain visual features; the visual features include camera movement patterns, scene switching frequency, density of main objects, visual rhythm, and dynamic changes;

[0069] The audio signal is determined based on the video content of the target video, and the Mel frequency cepstral coefficients are extracted from the audio signal. The short-time energy and zero-crossing rate of each audio frame in the audio signal are calculated, and the Mel frequency cepstral coefficients, each short-time energy, and the zero-crossing rate are used as audio features.

[0070] Text data is determined based on the video content of the target video, and the inverse document frequency weight of each semantic unit in the text data is calculated. The topic model is then performed based on the inverse document frequency weight of each word, and the topic distribution probability is obtained as the text feature.

[0071] As mentioned above, visual features are the feature information extracted from the video frames of the target video. Therefore, the video frames can be determined first based on the video content of the target video, and then a relevant network model (first network model) can be used to extract visual features. These visual features may include, but are not limited to, camera movement patterns, scene switching frequency, density of main objects, visual rhythm, and dynamic changes. In one possible implementation, the first network model can specifically be a 3D CNN (3D Convolutional Neural Network) model. Of course, the above visual feature extraction method is only one implementation method provided in this application and is not the only one; other methods can also be used. For example, a 2D CNN (2D Convolutional Neural Network) model can be used to extract feature information from a single frame of video, and then an RNN (Long Short-Term Memory) or LSTM (Long Short-Term Memory) can be used to perform temporal modeling of these feature information to achieve visual feature extraction. In addition, optical flow can be used to extract motion information in the target video as a supplement to visual features.

[0072] As mentioned above, audio features are the feature information extracted from the audio signal of the target video. Therefore, the audio signal can be extracted first from the video content of the target video, and then the Mel-Frequency Cepstral Coefficients (MFCC), short-time energy of each audio frame, and zero-crossing rate can be obtained as audio features. The short-time energy reflects the intensity change of the audio signal, and the zero-crossing rate reflects the frequency change and rhythmic characteristics of the audio signal. Of course, the above audio feature extraction method is only one implementation method provided in this application and is not the only one; other methods can also be used. For example, deep neural networks can be directly used to extract features from the audio signal to obtain audio features; or an audio event detection model can be directly used to detect specific events (such as laughter, applause, musical beats, etc.) in the audio signal as audio features.

[0073] As mentioned above, text features are the feature information extracted from the text data of the target video. Therefore, the text data can be determined first based on the video content of the target video. Then, by methods such as text segmentation, filtering stop words, and removing punctuation and special characters, each speech unit in the text data can be obtained. Finally, the word frequency inverse document frequency weight of each semantic unit is calculated for topic modeling, and the topic distribution probability is obtained as the text feature. The topic modeling process can be implemented using the Latent Dirichlet Allocation (LDA) algorithm. Of course, the above text feature extraction method is only one implementation method provided in this application and is not the only one. Other methods can also be used. For example, pre-trained word vector models (such as Word2Vec, GloVe, etc.) or language models (such as BERT, GPT, etc.) can also be used to extract text features.

[0074] S102: Perform sentiment analysis on the user comment data of the target video to obtain user sentiment characteristics; user sentiment characteristics include user sentiment distribution characteristics and user comment characteristics.

[0075] This step aims to extract user sentiment features from the target video. These features can be extracted from user comment data of the target video. The user sentiment features can include user sentiment distribution features and user comment features. User sentiment distribution features are the user sentiment distribution information in the user comment data, and user comment features are the feature information of the user comment data itself.

[0076] In one embodiment of this application, performing sentiment analysis on user comment data of the target video to obtain user sentiment characteristics may include:

[0077] The second network model is used to process the user comment data of the target video to obtain the sentiment type of each user comment data, and the user sentiment distribution characteristics are determined based on the sentiment type of each user comment data; the sentiment types include positive sentiment, negative sentiment, and neutral sentiment;

[0078] The total number of user comments and the growth rate of comment density are determined based on the user comment data of the target video, and the total number of user comments and the growth rate of comment density are used as user comment features.

[0079] As mentioned above, user sentiment distribution characteristics refer to the user sentiment distribution information in user comment data. Based on this, sentiment type identification can be performed on each user comment data to determine the sentiment type of each user comment data, thereby determining the user sentiment distribution characteristics. The sentiment type identification process can be implemented using a relevant network model (second network model). In one possible implementation, the second network model can specifically be a Bidirectional Long Short-Term Memory (BiLSTM) model. Of course, the above sentiment type identification method is only one implementation method provided in this application and is not the only one; other methods can also be used. For example, a sentiment dictionary can be used for rule matching to achieve sentiment type identification, or traditional sentiment classifiers (such as SVM, Naive Bayes, etc.) can be directly used to achieve sentiment type identification.

[0080] As mentioned above, user comment characteristics refer to the inherent features of user comment data. In this embodiment, user comment characteristics may include the total number of user comments and the comment density growth rate. The comment density growth rate can be calculated as the ratio of the comment increment within a sliding time window to the length of the time window, and this growth rate reflects the level of user activity in the discussion.

[0081] S103: Perform dynamic trajectory analysis on the user interaction behavior of the target video to obtain user behavior characteristics; user behavior characteristics include user behavior evolution characteristics.

[0082] This step aims to extract user behavior features from the target video. These features can be derived from user interaction behavior during video playback. These user behavior features can include user behavior evolution features, which characterize the evolutionary trajectory of user behavior from initial contact to deep interaction. User behavior may include, but is not limited to, the target video's play count (total number of times the video is played), likes (total number of times users like the video), favorites (total number of times users favorite the video), shares (total number of times users share the video), completion rate (percentage of users who watch the entire video), average viewing time (average viewing time per user), bounce rate (percentage of users who leave shortly after the video starts), and user dwell time distribution (distribution of user dwell time across different time periods of the video).

[0083] In one embodiment of this application, dynamic trajectory analysis of user interaction behavior in the target video is performed to obtain user behavior characteristics, which may include:

[0084] The user interaction behavior of the target video is sampled according to a preset time window to obtain the time sequence data of each user interaction behavior.

[0085] The third network model is used to model the temporal dependency relationship of each user interaction behavior time series data to obtain the user behavior evolution characteristics.

[0086] Specifically, user interaction behaviors in the target video can be sampled first according to a preset time window (e.g., a fixed time window of 720 minutes and a sampling step size of 60 minutes) to obtain multiple time-series data of user interaction behaviors (corresponding to the number of sampling times). Then, a relevant network model (the third network model) is used to model the temporal dependencies of each user interaction behavior time-series data to obtain the evolutionary characteristics of user behavior. In one possible implementation, the third network model can be a Gated Recurrent Unit (GRU) network model, an LSTM network model, or a Transformer network model.

[0087] S104: Perform cross-modal fusion processing on video features, user emotional features, and user behavioral features to obtain fused features.

[0088] This step aims to achieve cross-modal fusion processing of video features, user emotional features, and user behavioral features, so as to enable the popularity assessment of the target video based on the fused features.

[0089] In one embodiment of this application, cross-modal fusion processing of video features, user emotion features, and user behavior features to obtain fused features may include: using a multi-head self-attention network to calculate the weights of video features, user emotion features, and user behavior features to obtain attention weights for video features, user emotion features, and user behavior features; and using each attention weight to perform weighted fusion processing on the video features, user emotion features, and user behavior features to obtain fused features.

[0090] This application provides an implementation method for cross-modal feature fusion processing, namely, a multi-feature fusion method based on a multi-head self-attention network. In one possible implementation, the multi-head attention network can specifically be a four-head attention network. Of course, the above multi-feature fusion method is only one implementation method provided by this application and is not the only one; other methods can also be used. For example, feature vectors of various features can be directly concatenated to achieve multi-feature fusion, or a weighted fusion method based on a gating mechanism can be used to achieve multi-feature fusion.

[0091] S105: Determine the heat assessment results of the target video based on the fusion characteristics.

[0092] This step aims to achieve target video popularity assessment based on fusion features, that is, to determine the popularity assessment result of the target video based on fusion features that include multi-dimensional features (video features, user emotional features, user behavioral features).

[0093] Specifically, the popularity assessment result of the target video can be the popularity level of the target video. Based on this, in one embodiment of this application, determining the popularity assessment result of the target video according to the fusion features may include: processing the fusion features using a regressor network to obtain the popularity assessment value of the target video; determining the target popularity level to which the popularity assessment value belongs according to the popularity level classification rules, and using the target popularity level as the popularity assessment result of the target video.

[0094] Specifically, the regressor network can consist of three fully connected layers. It converts the fused features into continuous values ​​between 0 and 100, which serve as video popularity evaluation values. Then, it matches the popularity level classification rules to determine the target popularity level of the target video. For example, the popularity level classification rules may include: unpopular level: evaluation value < 30; ordinary level: 30 ≤ evaluation value < 60; popular level: 60 ≤ evaluation value < 90; viral level: evaluation value ≥ 90.

[0095] In one embodiment of this application, the video popularity assessment method may further include:

[0096] Calculate the time derivative of the heat assessment value;

[0097] An abnormal heat alarm is triggered when the absolute value of the time derivative exceeds a preset threshold.

[0098] Determine the rate of change in the sentiment distribution of user comments based on user comment data;

[0099] Determine the fluctuation pattern of user behavior metrics based on user interaction behavior.

[0100] The abnormal patterns corresponding to abnormal popularity alerts are determined and output based on the rate of change in the sentiment distribution of user comments and the fluctuation patterns of user behavior indicators.

[0101] The video popularity assessment method provided in this application embodiment can further implement popularity anomaly alarms and output the current anomaly mode of the target video. The popularity anomaly alarm can be determined by referring to the time derivative of the popularity assessment value: Time derivative = (Current video popularity assessment value - Previous video popularity assessment value) / Assessment time interval; in one possible implementation, the preset threshold can be set to 0.5. Based on this, the change rate of user comment sentiment distribution and the fluctuation mode of user behavior indicators can be determined by using user comment data and user interaction behavior collected during the two assessment processes of the target video, which can be used to determine the anomaly mode of the target video. The anomaly mode can include, but is not limited to, suspected fraudulent activity (e.g., the proportion of negative sentiment in the comment sentiment distribution increases by more than 20% in a short period of time and the number of views surges simultaneously), public opinion event-driven (e.g., the proportion of positive sentiment in the comment sentiment distribution increases by more than 30% and the growth rate of sharing exceeds the growth rate of viewing), platform recommendation intervention (e.g., abnormal growth rate of viewing but growth rate of likes and collections is lower than the threshold), etc.

[0102] Furthermore, it is understood that the various network models used in the above embodiments can be updated by periodically collecting new data from the target video, thereby effectively improving model accuracy and further enhancing the accuracy of video popularity assessment results.

[0103] As can be seen, the video popularity assessment method provided in this application can extract video features, user emotional features, and user behavioral features from a target video, and then fuse them to achieve video popularity assessment. This technical solution takes into account the implicit emotional feedback generated by users during video viewing, which helps improve the accuracy of video popularity assessment results. It achieves a dynamic, accurate, and interpretable assessment of the video's dissemination potential, providing more reliable data for subsequent content distribution, recommendation systems, and business decisions.

[0104] For example, this application provides a method for evaluating video popularity.

[0105] First, please refer to Figure 2 , Figure 2 This is a schematic diagram of the structure of a video popularity assessment system provided in an embodiment of this application. The video popularity assessment system mainly includes a video content feature extraction module, a user comment sentiment analysis module, a viewing behavior time sequence modeling module, a popularity comprehensive decision module, a popularity level classification module, an online model update mechanism, an abnormal popularity detection submodule, a batch parallel assessment architecture, and a visualization output module.

[0106] 1. Video Content Feature Extraction Module: Used to perform structured analysis on the input video and extract its visual features, audio features, and text features.

[0107] Please refer to Figure 3 , Figure 3 This is a schematic diagram of the structure of a video content feature extraction module provided in an embodiment of this application. The video content feature extraction module mainly includes a visual feature extraction submodule, an audio feature extraction submodule, a text feature extraction submodule, a feature normalization processing module, a horizontal splicing and fusion module, and an 1113-dimensional multimodal representation vector output. Based on this video content feature extraction module, the video content feature extraction process is as follows:

[0108] 1. Visual feature extraction:

[0109] (1) Video frame sequence preprocessing: The video is sampled at a sampling frequency of 25 frames per second. The size of each frame image is normalized (e.g., adjusted to 224×224 pixels) and data enhancement (e.g., random cropping, horizontal flipping) is performed.

[0110] (2) Feature encoding of three-dimensional convolutional neural network (3D CNN): The 3D CNN network model adopts a five-layer three-dimensional convolutional neural network structure. The kernel size of each layer is 3×3×3, the stride is 2, and the activation function is ReLU (Rectified Linear Unit). The first layer of convolution is used to capture local spatiotemporal features in the original pixel space. Subsequent layers gradually expand the receptive field through stacked convolution operations. Finally, the output is a high-dimensional vector with a dimension of 1024. This vector represents the abstract features of the video at the dynamic visual semantic level, covering lens motion mode, scene switching frequency, density of main object appearance, visual rhythm and dynamic changes, etc.

[0111] 2. Audio feature extraction:

[0112] (1) Audio signal preprocessing: The video audio signal is pre-emphasized to eliminate the high-frequency attenuation effect; Short-Time Fourier Transform (STFT) is performed with a window length of 25 milliseconds and a frame shift of 10 milliseconds.

[0113] (2) Mel frequency cepstral coefficient (MFCC) extraction: Extract the 13-dimensional Mel frequency cepstral coefficient sequence, which can effectively characterize the spectral features of audio and reflect the rhythm and pitch information of audio.

[0114] (3) Short-time energy and zero-crossing rate calculation: Calculate the short-time energy of each frame to reflect the intensity change of the audio; calculate the zero-crossing rate to reflect the frequency change and rhythmic characteristics of the audio; short-time energy and zero-crossing rate together characterize the emotional intensity and rhythm of the audio.

[0115] (4)Feature compression: Combine 13-dimensional MFCC, short-time energy, and zero-crossing rate to form an audio feature matrix, and then compress it into a 39-dimensional feature vector through global max pooling operation. This vector contains rhythm intensity and emotional fluctuation information.

[0116] 3. Text feature extraction:

[0117] (1)Text preprocessing: Perform unified encoding conversion (such as UTF-8 encoding) on the video title, tags, and subtitle content. After removing punctuation marks and special characters, use a dictionary-based word segmentation algorithm to segment the text. Finally, filter out stop words (such as meaningless words like "de", "le", "zai", etc.) and retain the core semantic units.

[0118] (2)Term Frequency-Inverse Document Frequency (TF-IDF) weighting: Calculate the TF-IDF weight of each semantic unit in the current video corpus: TF\text{-}IDF(t,d) = TF(t,d) \times IDF(t), where TF(t,d) represents the term frequency of term t in document d, and IDF(t) represents the inverse document frequency. TF-IDF can effectively highlight the keywords in the document and reduce the weight of common words.

[0119] (3)Topic modeling: Input the TF-IDF weights of each word into the Latent Dirichlet Allocation algorithm for topic modeling. The number of topics is fixed at 50, and a 50-dimensional topic distribution probability vector can be output. This vector represents the distribution density of the video content in the semantic topic space and reflects the topic tendency of the video.

[0120] 4. Feature fusion:

[0121] (1)Feature normalization: Perform zero-mean unit-variance normalization on the visual feature vector (1024-dimensional), audio feature vector (39-dimensional), and text feature vector (50-dimensional) respectively: x_{norm} = \frac{x - \mu}{\sigma}, where \mu is the mean and \sigma is the standard deviation.

[0122] (2)Feature concatenation: Horizontally concatenate on the feature dimension to form a 1113-dimensional (1024 + 39 + 50) multi-modal representation vector of video content. This vector serves as the input basis for subsequent cross-modal interaction modeling.

[0123] II. User comment sentiment analysis module: Quantitatively analyze the sentiment tendency of the user comment data accumulated after the video is released.

[0124] Please refer to Figure 4 , Figure 4 This is a schematic diagram of the structure of a user comment sentiment analysis module provided in an embodiment of this application. The module mainly includes a comment cleaning and preprocessing module, a bidirectional LSTM semantic encoding module, a sentiment polarity tri-classification module, a sentiment distribution histogram construction module, a discrete cosine transform compression module, a comment density growth rate calculation module, and a 5-dimensional user sentiment feedback feature vector output. Based on this module, the user comment sentiment analysis process is as follows:

[0125] 1. Comment preprocessing:

[0126] (1) Text cleaning: This may include, but is not limited to, unifying character encoding (such as converting to UTF-8), removing emojis and link addresses, standardizing uppercase and lowercase formats, and removing invalid comments with a length of less than 3 characters.

[0127] (2) Text standardization: handle online terms and abbreviations, and standardize the format of numbers and punctuation marks.

[0128] 2. Sentiment polarity analysis:

[0129] (1) Bidirectional Long Short-Term Memory (BiLSTM) Semantic Encoding: The cleaned comment text is input into a Bidirectional Long Short-Term Memory network for semantic encoding. The hidden layer dimension of the network is set to 256, and an 8-head attention mechanism is used to weight and aggregate the semantic context of the comment. It can be understood that Bidirectional LSTM can consider the context information of the comment at the same time, and better understand the semantics.

[0130] (2) Calculation of sentiment polarity score: Each comment will eventually output a sentiment polarity score between -1 and 1; a score close to 1 indicates positive sentiment, close to -1 indicates negative sentiment, and close to 0 indicates neutral sentiment.

[0131] 3. Sentiment Classification and Statistics:

[0132] (1) Three-classification of sentiment: The sentiment polarity score is divided into three categories according to the preset threshold: comments with a score greater than 0.3 are marked as positive sentiment; comments with a score less than -0.3 are marked as negative sentiment; and the rest of the comments (between -0.3 and 0.3) are marked as neutral sentiment.

[0133] (2) Construction of sentiment distribution histogram: The proportion of three types of sentiment comments in the total number of comments is statistically analyzed to form a three-dimensional sentiment distribution histogram: [positive proportion, neutral proportion, negative proportion]. This histogram reflects the user's overall sentiment tendency towards the video.

[0134] (3) Discrete Cosine Transform (DCT) compression: Perform discrete cosine transformation compression on the sentiment distribution histogram, retain the main sentiment distribution energy, and compress it into a low-dimensional feature vector with a dimension of 3.

[0135] 4. Comment density calculation:

[0136] (1) Total number of comments: The total number of comments recorded on the video.

[0137] (2) Calculation of comment density growth rate: The comment density growth rate reflects the level of user participation in the discussion. It can be calculated by the ratio of the comment increment within the sliding time window to the length of the time window, that is, comment density growth rate = \frac{\Delta C}{\Delta T}, where \Delta C is the comment increment within the time window and \Delta T is the length of the time window.

[0138] 5. Construction of user sentiment feedback feature vectors:

[0139] The emotional distribution compression coefficient (3-dimensional), the total number of comments (1-dimensional), and the comment density growth rate (1-dimensional) are combined to form a 5-dimensional user emotional feedback feature vector. This vector serves as a quantitative representation of the user's subjective value judgment and participates in subsequent popularity decisions.

[0140] III. Viewing Behavior Temporal Modeling Module: This module performs dynamic trajectory modeling of user interaction behavior during video playback.

[0141] Please refer to Figure 5 , Figure 5 This is a schematic diagram of a viewing behavior temporal modeling module provided in an embodiment of this application. The module mainly includes a multi-dimensional behavior indicator acquisition module, a sliding window sampling module, a standardization processing module, a gated recurrent unit temporal modeling module, a max-pooling feature extraction module, and a 128-dimensional behavior evolution feature vector output. Based on this viewing behavior temporal modeling module, the viewing behavior temporal modeling process is as follows:

[0142] 1. Behavioral metrics collection: The raw behavioral metrics collected by the system mainly include play count, like count, favorite count, share count, completion rate, average viewing time, bounce rate, and user dwell time distribution.

[0143] 2. Time-series data sequence construction:

[0144] (1) Sliding window sampling: Sliding sampling is performed with a fixed time window of 720 minutes (12 hours) and the sampling step size is set to 60 minutes (1 hour) to obtain a multidimensional time series data sequence, in which each time window contains all behavioral indicators within that time period.

[0145] (2) Data standardization: Standardize the behavioral data within each time window to eliminate differences in units and make different indicators comparable. The standardization formula is: x_{std} = \frac{x - \mu}{\sigma}, where \mu is the mean and \sigma is the standard deviation.

[0146] 3. Temporal dependency modeling:

[0147] (1) Gated Recurrent Unit (GRU) Network: The standardized time series data is input into the GRU network to model the time series dependency relationship. The hidden layer dimension of the network is set to 128, and a unidirectional structure is used to capture the unidirectional dependency relationship of the behavior index as it evolves over time.

[0148] (2) Hidden state sequence generation: The network outputs a hidden state vector for each time step. This vector represents the abstract features of the user behavior pattern within the current time window and can effectively capture the correlation and evolution trend between behavioral indicators.

[0149] (3) Feature extraction: Perform max pooling on the hidden state vectors of all time steps and extract the most representative behavioral evolution features. Finally, output a behavioral evolution feature vector with a dimension of 128.

[0150] 4. The significance of behavioral evolutionary characteristics:

[0151] Behavioral evolution feature vectors can effectively capture the evolutionary trajectory of user behavior from initial contact to deep interaction. For example, a slowdown in the growth rate of play counts accompanied by a sharp increase in the number of favorites may indicate that the content has the potential for long-tail dissemination; the coexistence of high completion rate and low sharing volume may reflect that the content has an immersive feel but lacks social dissemination motivation; the rapid increase in likes but the slow increase in the number of comments may reflect that the content resonates but has low discussion.

[0152] IV. Popularity Comprehensive Decision Module: This module integrates the above three types of feature vectors to generate the final video popularity assessment value.

[0153] Please refer to Figure 6 , Figure 6 This is a schematic diagram of the structure of a popularity comprehensive decision-making module provided in an embodiment of this application. The module mainly includes a video content multimodal representation vector, a user sentiment feedback feature vector, a behavioral evolution feature vector, an independent fully connected layer alignment module, a four-head self-attention cross-modal fusion module, a three-layer fully connected regressor, and a video popularity evaluation value output. Based on this popularity comprehensive decision-making module, the popularity comprehensive decision-making process is as follows:

[0154] 1. Feature space alignment:

[0155] The three types of feature vectors are input into independent fully connected layers for feature space alignment. Specifically, the video content multimodal representation vector (1113 dimensions) is mapped to a 256-dimensional intermediate representation vector through a 512-dimensional fully connected layer; the user emotion feedback feature vector (5 dimensions) is mapped to a 256-dimensional intermediate representation vector through a 256-dimensional fully connected layer; and the behavior evolution feature vector (128 dimensions) is mapped to a 256-dimensional intermediate representation vector through a 128-dimensional fully connected layer. The final output of all three is a unified intermediate representation vector with a dimension of 256.

[0156] 2. Cross-modal interaction modeling:

[0157] (1) Multi-head self-attention mechanism: A four-head self-attention mechanism is used to model the cross-modal interaction of three types of 256-dimensional intermediate representation vectors. Each attention head independently calculates the query, key, and value matrix, and generates the attention weight matrix through the scaling dot product attention mechanism. The attention weight reflects the importance of different modal features in the popularity assessment.

[0158] (2) Attention weight calculation: Based on the attention weight formula, Attention(Q,K,V) = softmax(\frac{QK^T}{\sqrt{d_k}})V, where d_k is the dimension of the key vector, and the scaling factor\sqrt{d_k} is used to prevent the gradient from vanishing due to the excessively large dot product value.

[0159] (3) Feature fusion: The three types of features are weighted and fused to finally output a joint feature vector with a dimension of 256. This vector integrates information from three dimensions: video content, user emotion, and viewing behavior.

[0160] 3. Generation of popularity assessment value:

[0161] (1) Regressor network: The joint feature vector input is a regressor consisting of three fully connected layers. The first layer has a dimension of 512, the second layer has a dimension of 256, and the third layer has a dimension of 128. The output layer uses a linear activation function to generate continuous values ​​between 0 and 100, which are used as video popularity evaluation values.

[0162] (2) Loss function design: The loss function of the regressor adopts a weighted combination of mean squared error (MSE) and ranking loss, with weight coefficients set to 0.7 and 0.3 respectively, i.e. Loss = 0.7 times MSE + 0.3 times RankingLoss; where mean squared error ensures the accuracy of absolute numerical prediction, and ranking loss ensures the accuracy of relative ranking.

[0163] V. Popularity Level Classification Module: Based on the preset popularity range threshold, the evaluation value is divided into different levels.

[0164] Please refer to Figure 7 , Figure 7 This is a schematic diagram of a popularity ranking module provided in an embodiment of this application. The module mainly includes a video popularity evaluation value input, a ranking module for niche / ordinary / popular / viral content, a recommendation weight coefficient mapping module, and an exposure strategy engine interface. Based on this popularity ranking module, the popularity ranking process is as follows:

[0165] 1. Grading rules:

[0166] Unpopularity level: Evaluation value < 30, corresponding recommendation weight coefficient 0.1;

[0167] Normal level: 30 ≤ evaluation value < 60, corresponding to a recommendation weight coefficient of 0.3;

[0168] Popularity level: 60 ≤ evaluation value < 90, corresponding to a recommendation weight coefficient of 0.6;

[0169] Bestseller rating: Evaluation score ≥ 90, corresponding recommendation weight coefficient 1.0.

[0170] 2. Application of Recommendation Weight: The recommendation weight coefficient directly affects the exposure strategy engine of the content distribution platform, determining the initial ranking position and exposure frequency of the video in the recommendation list; the popularity level classification results are simultaneously output to the operation monitoring system, allowing content creators and platform managers to review the dissemination effect and adjust strategies.

[0171] VI. Online Model Update Mechanism: During the continuous dissemination of the video, new data is collected periodically and the model parameters are updated.

[0172] Please refer to Figure 8 , Figure 8 This is a schematic diagram of the structure of an online model update mechanism provided in an embodiment of this application. The online model update mechanism mainly includes a module for collecting new comment and behavior data, a module for retaining historical data via a sliding window, a module for discarding expired data, a bidirectional LSTM incremental training module, a GRU incremental training module, a stochastic gradient descent optimizer module, a learning rate decay control module, and a parameter synchronization to inference nodes module. Based on this online model update mechanism, the online model update process is as follows:

[0173] 1. Incremental training mechanism:

[0174] (1) Data collection: Regularly collect new user comments and behavior data, set the update cycle to be synchronized with the video popularity evaluation cycle, and perform incremental training before each evaluation.

[0175] (2) Sliding window mechanism: Set to retain historical data within the most recent 30 time units and discard expired data beyond this time range to ensure that the model parameters always reflect the latest user preference distribution.

[0176] (3) Parameter update: The bidirectional long short-term memory network of the user comment sentiment analysis module and the gated recurrent unit network of the viewing behavior temporal modeling module both use the stochastic gradient descent (SGD) optimizer for parameter update. The initial learning rate is set to 0.001 and is adjusted at a decay rate of 0.95 after each training round.

[0177] (4) Parameter synchronization: The model parameter update process is executed on an independent computing node. After the update is completed, the new parameters are synchronized to the inference node to ensure the continuity and consistency of the evaluation process.

[0178] 2. Dynamic adaptability:

[0179] This mechanism enables the system to dynamically adapt to shifts in user preferences. For example, during specific holidays, users' emotional tendencies may be generally positive, and the system can automatically adjust the emotional polarity classification threshold to adapt to the new distribution pattern. As user behavior patterns change over time, the system can automatically learn new behavioral evolution patterns.

[0180] 7. Abnormal Popularity Detection Submodule: Calculates the time derivative of the video popularity assessment value in real time to identify non-natural traffic interference.

[0181] Please refer to Figure 9 , Figure 9 This is a schematic diagram of the structure of an abnormal popularity detection submodule provided in an embodiment of this application. The abnormal popularity detection submodule mainly includes a popularity assessment value time derivative calculation module, a mutation threshold judgment module, an abnormal popularity alarm triggering module, an emotion distribution change rate matching module, a behavioral indicator fluctuation pattern matching module, an abnormal pattern template library, and an abnormal log database writing module. Based on this abnormal popularity detection submodule, the abnormal popularity detection process is as follows:

[0182] 1. Calculation of time derivative:

[0183] (1) Derivative calculation: The time derivative is obtained by dividing the difference between the current evaluation value and the previous evaluation value by the evaluation time interval, that is, time derivative = \frac{current evaluation value - previous evaluation value}{evaluation time interval}.

[0184] (2) Mutation threshold judgment: If the absolute value of the time derivative exceeds the preset mutation threshold of 0.5, an abnormal heat alarm will be triggered. The threshold can be adjusted according to actual business needs.

[0185] (3) Abnormal heat alarm: Abnormal heat alarm includes video unique identifier, time point of mutation, absolute value of mutation magnitude, and possible cause analysis results.

[0186] 2. Analysis of possible contributing factors:

[0187] Based on the matching and inference of the change rate of sentiment distribution in user comments during the same period and the abnormal fluctuation patterns of behavioral indicators, the system presets three types of abnormal pattern templates:

[0188] (1) Suspected fraudulent activity: The proportion of negative emotions in the comment sentiment distribution increased by more than 20% in a short period of time and the number of views surged simultaneously, which is inferred to be suspected fraudulent activity.

[0189] (2) Driven by public opinion events: The proportion of positive emotions increased by more than 30% and the growth rate of sharing exceeded the growth rate of playback, which is inferred to be driven by public opinion events.

[0190] (3) Platform recommendation intervention: The abnormal growth rate of play volume but the growth rate of likes and collections is lower than the threshold, which is inferred to be platform recommendation intervention.

[0191] 3. Anomaly Handling: Anomaly detection results are not included in the final popularity value output, but are only written to the anomaly log database as an independent monitoring signal for the manual review team to intervene and investigate.

[0192] 8. Batch Parallel Evaluation Architecture: Used to support batch video parallel evaluation.

[0193] Please refer to Figure 10 , Figure 10 This is a schematic diagram of a batch parallel evaluation architecture provided in an embodiment of this application. The architecture mainly includes a video content feature extraction task, a user comment sentiment analysis task, a viewing behavior temporal modeling task, GPU cluster computing nodes, NLP dedicated accelerator nodes, CPU cluster computing nodes, a message queue asynchronous communication module, a distributed key-value database, and a consistent hashing sharding module. The architectural features of this batch parallel evaluation architecture are as follows:

[0194] 1. The video content feature extraction task is assigned to a GPU (Graphics Processing Unit) cluster to accelerate the process using parallel computing capabilities;

[0195] 2. User comment sentiment analysis tasks are assigned to a dedicated NLP (Natural Language Processing) accelerator to improve sentiment analysis efficiency;

[0196] 3. View the behavior time series modeling task and assign it to the CPU cluster to ensure the real-time performance of behavior modeling;

[0197] 4. Asynchronous communication between modules is achieved through message queues;

[0198] 5. Evaluation results are uniformly written to a distributed key-value database, supporting high-concurrency queries.

[0199] 9. Visualization Output Module: Used to present the evaluation results in a visual manner.

[0200] Please refer to Figure 11 , Figure 11 This is a schematic diagram of the structure of a visualization output module provided in an embodiment of this application. The visualization output module mainly includes video popularity assessment value, sentiment distribution histogram, behavior evolution trend curve, popularity level label, anomaly detection status, structured data package integration module, API interface output module, heat map / radar chart / line chart / bar chart front-end display module, and time axis backtracking function module.

[0201] Finally, the video popularity assessment system provided in this application embodiment can be applied to the following scenario embodiments.

[0202] Scenario Example 1: Short video platform video popularity assessment scenario.

[0203] 1. Scenario Description: On a short video platform, millions of new videos are uploaded every day. The platform needs to accurately assess the popularity of each video in order to optimize recommendation algorithms and traffic allocation.

[0204] 2. System operation process:

[0205] (1) Video content feature extraction: Input a 3-minute food preparation video, and perform visual feature extraction (3D CNN extracts 1024-dimensional visual features to capture visual elements such as close-ups of ingredients, cooking actions, and finished product display in the video), audio feature extraction (MFCC extracts 39-dimensional audio features to capture the rhythm and emotion of the background music), and text feature extraction (LDA extracts 50-dimensional topic distribution to identify topics such as "food", "tutorial", and "life"), and merge them into a 1113-dimensional multimodal representation vector.

[0206] (2) User comment sentiment analysis: 5,000 comments were accumulated within 24 hours after the video was released. The corresponding sentiment analysis results are as follows: 3,500 positive sentiment comments (70%); 1,000 neutral sentiment comments (20%); 500 negative sentiment comments (10%). Then the comment density growth rate was calculated: 200 comments per hour. Finally, a 5-dimensional user sentiment feedback feature vector [0.7, 0.2, 0.1, 5,000, 200] was generated.

[0207] (3) Temporal modeling of viewing behavior: Collect behavioral data within 12 hours, and slide sampling in 1-hour time windows to obtain the behavioral evolution trend: {First 3 hours: rapid increase in play volume and synchronous increase in like volume; 3-6 hours: slowdown in play volume growth, but rapid increase in collection volume; 6-12 hours: stable play volume and increase in sharing volume}. Then, use GRU network to capture the behavioral evolution pattern of "play → like → collection → share", and finally generate a 128-dimensional behavioral evolution feature vector.

[0208] (4) Popularity Comprehensive Decision: Input the three types of feature vectors into the popularity comprehensive decision module, and calculate the cross-modal attention weights by the multi-head self-attention mechanism: video content feature weight 0.4; user emotional feedback feature weight 0.35; behavioral evolution feature weight 0.25; the regressor outputs a popularity evaluation value of 75.3.

[0209] (5) Popularity level classification: The evaluation value of 75.3 belongs to the popular level (between 60 and 90), the recommendation weight coefficient is 0.6, and the exposure frequency of the video in the recommendation list increases by 60%.

[0210] (6) Online model update: After 24 hours, new comments and behavior data are collected, the data of the most recent 30 hours is retained using a sliding window mechanism, and the BiLSTM and GRU networks are incrementally trained; the updated model parameters are synchronized to the inference nodes.

[0211] 3. Comparison of effects:

[0212] (1) Traditional view count evaluation method: Only view count is considered. The video has a moderate view count and is rated as ordinary. The recommendation weight is 0.3.

[0213] (2) Sentiment analysis evaluation method of this scheme: taking into account content quality, user sentiment and behavioral evolution, it is evaluated as popular level, with a recommendation weight of 0.6.

[0214] The comparison shows that the actual results of this solution are: a 25% increase in user satisfaction and a 30% increase in repeat visit rate.

[0215] Scenario Example 2: Accurate identification of controversial videos.

[0216] 1. Scenario Description: A video has sparked considerable discussion due to its controversial content, resulting in high view counts, but user sentiment is complex. It's crucial to accurately identify this situation to avoid misjudging controversial content as high-quality content.

[0217] 2. System operation process:

[0218] (1) Video content feature extraction: Input a 5-minute video discussing a social topic and extract a 1113-dimensional multimodal representation vector.

[0219] (2) User comment sentiment analysis: 20,000 comments were accumulated within 48 hours after the video was released. The corresponding sentiment analysis results were: 6,000 positive sentiment comments (30%); 4,000 neutral sentiment comments (20%); and 10,000 negative sentiment comments (50%). The comment density growth rate was then calculated to be 500 comments per hour (abnormally high). Finally, a 5-dimensional user sentiment feedback feature vector was generated: [0.3, 0.2, 0.5, 20,000, 500].

[0220] (3) Temporal modeling of viewing behavior: Collect behavioral data within 12 hours, slide sampling in 1-hour time windows to obtain behavioral evolution trends: {rapid growth in play count; slow growth in like count; almost no growth in collection count; abnormally high sharing count (user forwarding and discussion)}, and then use GRU network to identify abnormal behavior patterns, and finally generate a 128-dimensional behavioral evolution feature vector.

[0221] (4) Popularity comprehensive decision: Although the number of views is high, the proportion of negative emotions is high (50%) and the behavior pattern is abnormal. The weight of user emotional feedback features is reduced by the multi-head self-attention mechanism. Finally, the regressor outputs a popularity evaluation value of 45.2 (ordinary level, not popular).

[0222] (5) Abnormal popularity detection: When the time derivative of the popularity assessment value is 0.8 (exceeding the threshold of 0.5), an abnormal popularity alarm is triggered. Among them, the cause analysis is that the proportion of negative emotions increases by 20% and the number of views surges, which is inferred to be suspected of being a traffic fraud or public opinion event. Finally, the abnormal alarm is written to the log for manual review.

[0223] 3. Comparison of effects:

[0224] (1) Traditional view count evaluation method: high view count is evaluated as popular level, with a recommendation weight of 0.6, resulting in a large number of controversial content being recommended;

[0225] (2) Sentiment analysis assessment method of this scheme: if the proportion of negative emotions is high, the assessment is ordinary level and the recommendation weight is 0.3 to avoid controversial content being misjudged.

[0226] The comparison shows that the actual effect of this solution is: the platform user reporting rate decreased by 40%, and the community atmosphere was significantly improved.

[0227] Scenario Example 3: Capturing the Synergistic Effect of Multimodal Emotions.

[0228] 1. Scene Description: A video successfully evoked sadness in viewers through the coordinated use of background music, visual color scheme, and subtitles. It is necessary to accurately capture this cross-modal emotional synergy.

[0229] 2. System operation process:

[0230] (1) Video content feature extraction: Input a touching story video with a duration of 4 minutes, and extract visual features (3D CNN captures that the screen tone is cool, mainly blue, and the scene transitions are slow), audio features (MFCC captures that the background music has a slow rhythm and a low tone), and text features (LDA identifies themes such as "emotion", "memories", and "farewell"), and fuse them into a 1113-dimensional multimodal representation vector. Among them, when the three types of features are analyzed separately, the emotional intensity is moderate, but through multimodal fusion, the system identifies an emotional co-reinforcement effect.

[0231] (2) User comment sentiment analysis: 8,000 comments were accumulated within 36 hours after the video was released. The corresponding sentiment analysis results are as follows: 7,200 positive sentiment comments (90%); 600 neutral sentiment comments (7.5%); 200 negative sentiment comments (2.5%). A large number of emotional words such as "moved", "tearful", and "resonance" appeared in the comments. Finally, a 5-dimensional user sentiment feedback feature vector was generated: [0.9, 0.075, 0.025, 8000, 220].

[0232] (3) Temporal modeling of viewing behavior: Collect behavioral data within 12 hours, and slide sampling in 1-hour time windows to obtain the behavioral evolution trend: {Stable growth in play count; abnormally high completion rate (85%, far exceeding the average level of 50%); rapid growth in collection count (users want to watch again); high sharing count (users want to share with friends)}. Then, use the GRU network to identify the high-quality content pattern of "high completion rate + high collection rate + high sharing rate", and finally generate a 128-dimensional behavioral evolution feature vector.

[0233] (4) Popularity comprehensive decision: Based on the multi-head self-attention mechanism, the emotional synergistic reinforcement effect of visual, audio and text features is identified; user emotional feedback is highly positive (90% positive); behavioral patterns show high content quality (high completion rate, high collection rate); the regressor outputs a popularity evaluation value of 88.5 (close to the level of a hit).

[0234] (5) The popularity level assessment value of 88.5 belongs to the popular level (between 60 and 90), which is close to the best-selling item. The recommendation weight coefficient is 0.6. The system suggests increasing the recommendation of similar content to extend the user's stay time.

[0235] 3. Comparison of effects:

[0236] (1) Traditional single-modal analysis method: only analyzes visual or audio features, cannot capture synergistic effects, and the evaluation value is too low;

[0237] (2) Multimodal fusion method of this scheme: accurately capture cross-modal emotional synergy effect, and the evaluation value accurately reflects the content quality.

[0238] As can be seen from the comparison, the actual effect of this solution is: user completion rate increased by 35%, and user dwell time increased by 40%.

[0239] In summary, the video popularity assessment method provided in this application has the following technical advantages:

[0240] 1. Integrating user emotional feedback significantly improves the accuracy of assessments and the alignment with user satisfaction: By introducing a user comment sentiment analysis module, the system can quantify users' emotional tendencies and distribution, enabling assessment results to reflect users' subjective value judgments. By establishing a quantitative mapping relationship between emotional intensity and dissemination momentum, the system can distinguish the essential differences between "highly controversial" and "highly positive" sentiments, significantly improving the alignment between assessment results and user satisfaction.

[0241] 2. Multimodal Feature Fusion and Cross-Modal Interaction Modeling to Capture Emotional Synergistic Amplification Effects: By extracting visual, audio, and textual multimodal features through a video content feature extraction module, and employing a multi-head self-attention mechanism for cross-modal interaction modeling, the synergistic amplification effects between emotional elements can be captured. This ensures that the evaluation results are highly consistent with real user psychological experiences, enabling an accurate understanding of cross-modal emotional coupling characteristics.

[0242] 3. Temporal modeling and online update mechanism for dynamic adaptability: By capturing the evolutionary trajectory of user behavior through the behavioral temporal modeling module, and dynamically adapting to the latest user preference distribution through the online model update mechanism, the system possesses the ability to dynamically adapt to user preference drift, ensuring the timeliness and forward-looking nature of the evaluation results.

[0243] 4. Anomaly detection mechanism enhances the evaluation system's resilience: The anomaly detection submodule calculates the time derivative of the popularity assessment value in real time, identifying non-natural traffic interference. Based on the rate of change in sentiment distribution and the fluctuation patterns of behavioral indicators, it infers the underlying causes, effectively identifying anomalies such as fraudulent traffic, public opinion events, and platform recommendation interventions, thus enhancing the evaluation system's resilience and credibility.

[0244] 5. The overall solution boasts high scalability and engineering feasibility: Employing a distributed computing architecture, different tasks are allocated to GPU clusters, dedicated NLP accelerators, and CPU clusters, with asynchronous communication via message queues supporting batch video parallel evaluation. This architecture ensures the system maintains evaluation timeliness even under high concurrency scenarios, meeting the real-time recommendation needs of large-scale content distribution platforms.

[0245] This application provides a video popularity assessment device.

[0246] Please refer to Figure 12 , Figure 12This is a schematic diagram of the structure of a video popularity assessment device provided in an embodiment of this application. The video popularity assessment device may include:

[0247] The first processing module 1 is used to extract features from the video content of the target video to obtain video features; the video features include visual features, audio features, and text features;

[0248] The second processing module 2 is used to perform sentiment analysis on the user comment data of the target video to obtain user sentiment features; the user sentiment features include user sentiment distribution features and user comment features.

[0249] The third processing module 3 is used to perform dynamic trajectory analysis on the user interaction behavior of the target video to obtain user behavior characteristics; the user behavior characteristics include user behavior evolution characteristics.

[0250] Fusion module 4 is used to perform cross-modal fusion processing on video features, user emotional features, and user behavior features to obtain fused features;

[0251] Module 5 is used to determine the heat assessment result of the target video based on the fusion characteristics.

[0252] As can be seen, the video popularity assessment device provided in this application embodiment can extract video features, user emotional features, and user behavioral features of a target video, and then perform fusion processing to achieve video popularity assessment of the target video. It is evident that this technical solution takes into account the implicit emotional feedback generated by users during video viewing to achieve video popularity assessment, which helps to improve the accuracy of video popularity assessment results. It realizes a dynamic, accurate, and interpretable assessment of the video's dissemination potential, and can provide more reliable data basis for subsequent content distribution, recommendation systems, and business decisions.

[0253] In one embodiment of this application, the first processing module 1 can be specifically used to determine video frames based on the video content of the target video, and process the video frames using a first network model to obtain visual features; the visual features include camera movement patterns, scene switching frequency, density of main objects, visual rhythm, and dynamic changes; determine audio signals based on the video content of the target video, extract Mel frequency cepstral coefficients from the audio signals, calculate the short-time energy and zero-crossing rate of each audio frame in the audio signals, and use the Mel frequency cepstral coefficients, each short-time energy, and the zero-crossing rate as audio features; determine text data based on the video content of the target video, calculate the word frequency inverse document frequency weights of each semantic unit in the text data, perform topic modeling based on each word frequency inverse document frequency weight, and obtain the topic distribution probability as text features.

[0254] In one embodiment of this application, the second processing module 2 can be specifically used to process the user comment data of the target video using a second network model to obtain the sentiment type of each user comment data, and determine the user sentiment distribution characteristics based on the sentiment type of each user comment data; the sentiment type includes positive sentiment, negative sentiment, and neutral sentiment; the total number of user comments and the comment density growth rate are determined based on the user comment data of the target video, and the total number of user comments and the comment density growth rate are used as user comment characteristics.

[0255] In one embodiment of this application, the third processing module 3 can be specifically used to sample the user interaction behavior of the target video according to a preset time window to obtain the time series data of each user interaction behavior; and to use a third network model to model the time series data of each user interaction behavior to obtain the user behavior evolution characteristics.

[0256] In one embodiment of this application, the fusion module 4 can be specifically used to calculate the weights of video features, user emotion features, and user behavior features using a multi-head self-attention network to obtain the attention weights of video features, user emotion features, and user behavior features; and to perform weighted fusion processing on the video features, user emotion features, and user behavior features using each attention weight to obtain fused features.

[0257] In one embodiment of this application, the determination module 5 can be specifically used to process the fusion features using a regressor network to obtain the popularity evaluation value of the target video; determine the target popularity level to which the popularity evaluation value belongs according to the popularity level classification rules, and use the target popularity level as the popularity evaluation result of the target video.

[0258] In one embodiment of this application, the video popularity assessment device may further include an alarm module for calculating the time derivative of the popularity assessment value; triggering an abnormal popularity alarm when the absolute value of the time derivative exceeds a preset threshold; determining the rate of change of user comment sentiment distribution based on user comment data; determining the fluctuation pattern of user behavior indicators based on user interaction behavior; and determining and outputting the abnormal pattern corresponding to the abnormal popularity alarm based on the rate of change of user comment sentiment distribution and the fluctuation pattern of user behavior indicators.

[0259] For a description of the apparatus provided in the embodiments of this application, please refer to the above method embodiments; further details will not be repeated here.

[0260] This application provides an electronic device.

[0261] Please refer to Figure 13 , Figure 13 This application provides a schematic diagram of the structure of an electronic device, which may include:

[0262] Memory 11 is used to store computer programs;

[0263] The processor 10 is configured to execute computer programs to implement the steps of any of the video heat assessment methods described above.

[0264] like Figure 13 The diagram shows the structural composition of an electronic device, which may include a processor 10, a memory 11, a communication interface 12, and a communication bus 13. The processor 10, memory 11, and communication interface 12 all communicate with each other through the communication bus 13.

[0265] In this embodiment, the processor 10 may be a central processing unit (CPU), an application-specific integrated circuit, a digital signal processor, a field-programmable gate array, or other programmable logic devices.

[0266] The processor 10 can call the program stored in the memory 11. Specifically, the processor 10 can execute the operations in the embodiments of the video heat assessment method.

[0267] The memory 11 is used to store one or more programs. The programs may include program code, which includes computer operation instructions. In this embodiment, the memory 11 stores at least a program for implementing the following functions:

[0268] Feature extraction is performed on the target video content to obtain video features, including visual, audio, and text features. Sentiment analysis is performed on user comment data of the target video to obtain user sentiment features, including user sentiment distribution features and user comment features. Dynamic trajectory analysis is performed on user interaction behavior of the target video to obtain user behavior features, including user behavior evolution features. Cross-modal fusion processing is performed on video features, user sentiment features, and user behavior features to obtain fused features. The popularity assessment result of the target video is determined based on the fused features.

[0269] In one possible implementation, the memory 11 may include a program storage area and a data storage area, wherein the program storage area may store the operating system and applications required for at least one function; and the data storage area may store data created during use.

[0270] In addition, memory 11 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device or other volatile solid-state storage device.

[0271] Communication interface 12 can be an interface for the communication module, used to connect with other devices or systems.

[0272] Of course, it should be noted that, Figure 13 The structure shown does not constitute a limitation on the electronic device in the embodiments of this application. In practical applications, the electronic device may include more than Figure 13 More or fewer components as shown, or combinations of certain components.

[0273] This application provides a computer-readable storage medium.

[0274] The computer-readable storage medium provided in this application embodiment stores a computer program, which, when executed by a processor, can implement the steps of any of the video heat assessment methods described above.

[0275] The computer-readable storage medium may include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0276] For a description of the computer-readable storage medium provided in this application, please refer to the above method embodiments; further details will not be repeated here.

[0277] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0278] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0279] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0280] The technical solutions provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only for the purpose of helping to understand the methods and core ideas of this application. It should be noted that those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A method for evaluating video popularity, characterized in that, include: Feature extraction is performed on the video content of the target video to obtain video features; the video features include visual features, audio features, and text features; Sentiment analysis is performed on the user comment data of the target video to obtain user sentiment features; the user sentiment features include user sentiment distribution features and user comment features. Dynamic trajectory analysis is performed on the user interaction behavior of the target video to obtain user behavior characteristics; The user behavior characteristics include user behavior evolution characteristics; The video features, user emotion features, and user behavior features are subjected to cross-modal fusion processing to obtain fused features; The heat assessment result of the target video is determined based on the fusion features.

2. The video popularity assessment method according to claim 1, characterized in that, Feature extraction is performed on the video content of the target video to obtain video features, including: Based on the video content of the target video, video frames are determined, and the video frames are processed using a first network model to obtain the visual features; the visual features include camera movement patterns, scene switching frequency, density of main objects, visual rhythm, and dynamic changes; The audio signal is determined based on the video content of the target video, and the Mel frequency cepstral coefficients are extracted from the audio signal. The short-time energy and zero-crossing rate of each audio frame in the audio signal are calculated, and the Mel frequency cepstral coefficients, each short-time energy, and the zero-crossing rate are used as the audio features. Text data is determined based on the video content of the target video, and the inverse document frequency weight of each semantic unit in the text data is calculated. Topic modeling is performed based on each inverse document frequency weight to obtain the topic distribution probability as the text feature.

3. The video popularity assessment method according to claim 1, characterized in that, Sentiment analysis is performed on user comment data for the target video to obtain user sentiment characteristics, including: The second network model is used to process the user comment data of the target video to obtain the sentiment type of each user comment data, and the user sentiment distribution characteristics are determined based on the sentiment type of each user comment data; the sentiment type includes positive sentiment, negative sentiment, and neutral sentiment; The total number of user comments and the comment density growth rate are determined based on the user comment data of the target video, and the total number of user comments and the comment density growth rate are used as the user comment features.

4. The video popularity assessment method according to claim 1, characterized in that, Dynamic trajectory analysis is performed on the user interaction behavior of the target video to obtain user behavior characteristics, including: The user interaction behavior of the target video is sampled according to a preset time window to obtain the time sequence data of each user interaction behavior. The user behavior evolution characteristics are obtained by modeling the temporal dependencies of the time-series data of each user interaction behavior using a third network model.

5. The video popularity assessment method according to claim 1, characterized in that, The video features, user sentiment features, and user behavior features are subjected to cross-modal fusion processing to obtain fused features, including: A multi-head self-attention network is used to calculate the weights of the video features, the user emotion features, and the user behavior features to obtain the attention weights of the video features, the user emotion features, and the user behavior features. The video features, user emotion features, and user behavior features are weighted and fused using the attention weights to obtain the fused features.

6. The video popularity assessment method according to claim 1, characterized in that, The heat assessment result of the target video is determined based on the fusion features, including: The fused features are processed using a regressor network to obtain the heat evaluation value of the target video; The target popularity level to which the popularity assessment value belongs is determined according to the popularity level classification rules, and the target popularity level is used as the popularity assessment result of the target video.

7. The video popularity assessment method according to claim 6, characterized in that, Also includes: Calculate the time derivative of the heat assessment value; When the absolute value of the time derivative exceeds a preset threshold, an abnormal heat alarm is triggered. The rate of change in the sentiment distribution of user comments is determined based on the user comment data. The fluctuation pattern of user behavior indicators is determined based on the user interaction behavior. The abnormal pattern corresponding to the abnormal popularity alarm is determined and output based on the change rate of the user comment sentiment distribution and the fluctuation pattern of the user behavior indicators.

8. A video popularity assessment device, characterized in that, include: The first processing module is used to extract features from the video content of the target video to obtain video features; the video features include visual features, audio features, and text features. The second processing module is used to perform sentiment analysis on the user comment data of the target video to obtain user sentiment features; the user sentiment features include user sentiment distribution features and user comment features. The third processing module is used to perform dynamic trajectory analysis on the user interaction behavior of the target video to obtain user behavior characteristics. The user behavior characteristics include user behavior evolution characteristics; The fusion module is used to perform cross-modal fusion processing on the video features, the user emotion features, and the user behavior features to obtain fused features; A determination module is used to determine the heat assessment result of the target video based on the fusion features.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the steps of the video heat assessment method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the video heat assessment method as described in any one of claims 1 to 7.

Images