A priori prediction method for video popularity integrating video text content

Abstract

The present invention relates to the field of computer application technology, and to a method for a priori prediction of video popularity integrating video text content; the method comprises the following steps: obtaining video data, extracting video text content, and classifying popularity according to the amount of playback; extracting keywords and vectorizing them; calculating the similarity between the word vector of a video to be predicted published by a current author and the word vector of videos published by other authors; selecting K videos with the highest ranking as first related videos, selecting Z videos recently published by the author of the video to be predicted as second related videos, and calculating the mean of corresponding attribute popularity numerical features respectively; averaging the mean of the popularity numerical features of the related videos again to generate attribute features; classifying the attribute features of videos related to the video to be predicted and the corresponding popularity, and predicting the a priori popularity of the video to be predicted by using a random forest algorithm based on a decision tree.

Description

Technical Field

[0001] The present invention relates to the field of computer application technology, and in particular to a method for a priori prediction of video popularity integrating video text content. Background Art

[0002] At present, with the development of communication technology and the popularization of 5G, the proportion of video services is increasing, and the proportion of mobile traffic generated by video in global Internet traffic is increasing. However, the caching capacity of edge cache facilities is limited and can only be selectively cached. Obviously, caching devices can more efficiently utilize cache resources by caching videos with higher popularity. In addition, studies have shown that the distribution of video popularity is extremely uneven, which is reflected in the fact that a few videos have a large number of views, while most videos have very few views, which shows the necessity and importance of video popularity prediction.

[0003] The current method mainly makes predictions based on the historical playback data of the video. However, this method is not applicable to newly uploaded videos and lacks mining of video text content. Therefore, a method is needed that integrates video text content and can make predictions for newly uploaded videos. Summary of the invention

[0004] The present invention designs a method for predicting video popularity a priori by integrating video text content, realizes the prediction of newly uploaded videos, and simultaneously mines the video text content. The method comprises:

[0005] A method for predicting video popularity a priori by integrating video text content, which is characterized by comprising the following steps:

[0006] Obtain a large amount of video data from the target video website, including text features and popularity numerical features of different attributes, to form a structured data set;

[0007] Preprocess the video data, extract the video text content, sort the video data in descending order according to the number of views, and mark the popularity classification of the video data;

[0008] The preprocessed video text content is extracted using the TextRank algorithm, and the extracted keywords are vectorized using the word2vec algorithm;

[0009] The cumulative calculation method is used for multiple keywords of each video data to calculate the similarity between the word vector of the video to be predicted published by the current author and the word vector of videos published by other authors;

[0010] Select K other videos with the highest similarity ranking as the first related videos of the video to be predicted, and calculate the mean of the corresponding attribute popularity numerical features of the K related videos respectively;

[0011] Select the Z videos recently released by the author of the video to be predicted as the second related videos, and calculate the mean of the corresponding attribute popularity numerical features of the Z related videos respectively;

[0012] The average of the popularity numerical features of the first related video and the second related video corresponding to the attribute of the video to be predicted is calculated again to generate an attribute feature;

[0013] The attribute features of videos related to the video to be predicted and their corresponding popularity are classified, and the prior popularity of the video to be predicted is predicted using the random forest algorithm based on decision trees.

[0014] Beneficial effects of the present invention:

[0015] The present invention can get rid of the dependence on historical data when predicting by starting from the perspective of video text content and the dimension of video author. In addition, the present invention expands the diversity of related videos by comparing the similarities between videos in the form of word vectors, making the features of the video to be predicted richer. By using decision-making to build a prediction algorithm, the complexity of the prediction algorithm is reduced, and it is easy to use large-scale data sets for training and rapid prediction deployment. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a framework diagram of a priori prediction algorithm for video popularity integrating video text content in an embodiment of the present invention;

[0017] Figure 2 The flowchart of the method for a priori prediction of video popularity integrating video text content in an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0018] The following is a further description of the specific implementation of the present invention in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention but are not intended to limit the scope of the present invention.

[0019] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] The overall framework of the video popularity prior prediction algorithm constructed by the present invention is shown in Figure 1As shown, it mainly includes data collection and preprocessing, related video determination, and classification algorithm construction prediction. First, data collection mainly includes crawling and parsing the video web page data of the target video website to obtain the original data of the video. The original video mainly includes text information and video information, that is, including title, tag, introduction description, video playback volume, number of likes, number of shares and other numerical features; video data preprocessing includes text data cleaning, such as removing pause words, punctuation marks, expressions and other interference factors, and marking the popularity classification of the video at the same time. Then, the related videos of the video are found from the two dimensions of keyword relevance and author relevance. Specifically, the keyword relevance dimension is to extract keywords from the video text data and vectorize the keyword words, calculate the correlation between the word vectors of the keywords and sort and select the appropriate number of related videos; the author relevance dimension is directly selected from the video recently released by the author. Finally, considering that the video has multiple features and limited popularity types, the present invention uses a classification algorithm to predict the popularity of the video, that is, to predict the popularity classification of the video to be uploaded by the mean of the related video features of the video to be uploaded. The algorithms used in the experiment include decision trees and random forests. The core idea of ​​the random forest algorithm is to build an integrated classifier using multiple decision trees as basic classifiers. One decision tree obtains one predicted classification result, and multiple decision trees obtain multiple different results. The classification with the largest number of categories is counted and used as the prediction result of the video to be predicted.

[0021] Figure 2 is a flow chart of a priori prediction of video popularity integrating video text content in an embodiment of the present invention, such as Figure 2 As shown, the method includes:

[0022] 101. Obtain a large amount of video data from the target video website, including text features and popularity numerical features of different attributes, to form a structured data set;

[0023] In an embodiment of the present invention, a large amount of video data can be obtained from a target video website through crawler technology. When obtaining the video data, the numerical features and text features of the video popularity are mainly crawled, where the numerical features include the number of views, the number of comments, the number of reposts, the number of favorites, the number of fans of the publishing author, etc., which can quantify the popularity of the video. The text features include the video title, video tags, video introduction, and the author's accompanying text description of the video, forming a structured data set.

[0024] It can be understood that, in the embodiment of the present invention, the target video website can be any public video website, such as iQiyi, Bilibili Video Network, Youku, YouTube, etc. Since each target video website has a different operating model, different target video websites have popularity value characteristics of corresponding attributes. Taking Bilibili Video Network as an example, the popularity value characteristics of different attributes include the number of followers of the uploader (Followers Count, FO), the number of views (View Count, VC), the number of comments (Comment Count, CM), the number of likes (Like Count, LC), the number of favorites (Favorite Count, FC), the number of shares (Share Count, SC), and the number of coins (Coins Count, CI).

[0025] 102. Preprocess the video data, extract the video text content, sort the video data in descending order according to the amount of playback, and mark the popularity classification of the video data;

[0026] In the embodiment of the present invention, the video data needs to be preprocessed, and the text feature operations include removing pause words, punctuation marks, numbers, and expressions to form a structured database; the video is classified into high, medium, and low according to the amount of video data played, and the marking is based on high:[c [0.1N] ,c1],medium:[c [0.4N] ,c [0.1N]+1 ],low:[c N ,c [0.4N]+1 ], where c i Indicates the number of views of the video ranked i, i∈[1,2,...,N], [0.1N] indicates the maximum integer value not exceeding 0.1N, [0.4N] indicates the maximum integer value not exceeding 0.4N, [N] indicates the maximum integer value not exceeding N, N indicates the total number of videos, and the subscript [] indicates rounding of numbers.

[0027] It is understandable that the lower right corners of the above categories are ranking orders, and the rankings are discrete integers. i Indicates the number of views of the video ranked i, and the smaller the ranking value, the larger the number of views. Therefore, the interval of the value is small on the left and large on the right. The interval is a closed interval with values ​​on both sides, represented by []. Therefore, different closed intervals need to be processed by +1 at the critical point. The meaning of adding 1 to the ranking value is that the ranking of the view volume drops by one, that is, C i >C i+1 , where [0.1N] is a common rounding operation in mathematics, for example, 0.1N=19.2, then [0.1N]=19.

[0028] Since users need to click on the video to interact with other related values, it is a subjective decision made by the users watching the video, and the number of views will automatically increase when the user clicks on the video, which shows that the view count feature is more reasonable for measuring video popularity.

[0029] 103. Extract keywords from the preprocessed video text content using the TextRank algorithm, and perform word vectorization processing on the extracted keywords using the word2vec algorithm;

[0030] The TextRank algorithm is a graph-based ranking algorithm for keyword extraction and document summarization. It is improved from Google's PageRank algorithm, which is a web page importance ranking algorithm. It can extract keywords by using the co-occurrence information (semantics) between words within a document. It can extract keywords and keyword groups from a given text, and use an extractive automatic abstract method to extract the key sentences of the text. Word2vec is an NLP tool launched by Google in 2013. Its feature is to vectorize all words. In this way, the relationship between words can be quantitatively measured, and the connection between words can be mined. Therefore, the embodiment of the present invention uses the TextRank algorithm to extract keywords from the video text content, and uses the word2vec algorithm to vectorize the keywords extracted by the TextRank algorithm to obtain keyword word vectors; wherein, the keywords of the video text content are obtained from the long text formed by the video title, video tags, video introduction, and the author's accompanying text description of the video.

[0031] 104. Using a cumulative calculation method for multiple keywords of each video data, calculate the similarity between the word vector of the video to be predicted published by the current author and the word vector of the video published by other authors;

[0032] In an embodiment of the present invention, the present invention respectively calculates the similarities between the M keyword word vectors of the video to be predicted and the N keyword word vectors of videos published by other authors, accumulates the similarities of the M keyword word vectors, and calculates the similarity between the word vector of the video to be predicted published by the current author and the word vector of videos published by other authors. For example, assuming that there are 6 keyword word vectors in the video to be predicted, and video A published by other authors has 5 keyword word vectors, video B has 6 keyword word vectors, and video C has 7 keyword word vectors; since the present invention adopts an accumulation method to calculate the similarity distance of the word vector, that is, the 6 keyword word vectors of the video to be predicted are respectively compared with the keyword word vectors of videos A, B and C for similarity, if there is no corresponding keyword of the video to be predicted in video A, then the distance is regarded as 0. According to the above method, the sum of the similarity distances between the 6 keywords of the video to be predicted and videos A, B and C are respectively calculated, and the sum of the distances is used as the final similarity between the word vector of the video to be predicted published by the current author and the word vector of videos published by other authors.

[0033] Prediction is performed by collecting numerical features related to the video. The determination of related videos is mainly from two aspects. First, through author relevance, specifically selecting the author's most recently released Z videos. Second, using keyword relevance to select K videos. Keyword relevance is mainly calculated by calculating the similarity of the word vectors of the video to be predicted and other videos in the video data. The calculation formula is:

[0034]

[0035] Among them, key1 represents the keyword of the video to be predicted, vec1 is the keyword word vector corresponding to key1, key2 represents the keyword of the video released by other authors, and vec2 is the keyword word vector corresponding to key2; is the value of the hth dimension of the vec1 vector, It is the value of the h-th dimension of the vec2 vector, where H represents the word vector dimension.

[0036] 105. Select K other videos with the highest similarity ranking as first related videos of the video to be predicted, and calculate the mean values ​​of the corresponding attribute popularity numerical features of the K related videos respectively;

[0037] In an embodiment of the present invention, the related videos of the video to be predicted published by the current author are divided into two categories, one is the videos published by other users, and the other is the videos recently published by the current author; therefore, the videos with high similarity ranking among the videos published by other users are taken as the first related videos of the video to be predicted, and the mean of each attribute popularity feature in the first related video is calculated respectively.

[0038] 106. Select Z videos recently released by the author of the video to be predicted as second related videos, and calculate the mean values ​​of corresponding attribute popularity numerical features of the Z related videos respectively;

[0039] In the embodiment of the present invention, the video most recently released by the current author is directly used as the second related video, and the mean value of each attribute popularity feature in the second related video is calculated respectively.

[0040] 107. averaging the averages of the popularity numerical features of the first related video and the second related video corresponding to the attribute of the video to be predicted, and generating attribute features;

[0041] In the embodiment of the present invention, the attribute means of the two related videos are comprehensively considered, and the corresponding attribute means of the two related videos are further averaged to obtain the attribute characteristics of the attribute.

[0042] 108. Classify the attribute features and corresponding popularity of videos related to the video to be predicted, and use the random forest algorithm based on the decision tree to predict the prior popularity of the video to be predicted.

[0043] In an embodiment of the present invention, the present invention sorts the attribute features of each attribute of the related videos of the to-be-predicted video, finds the best classification division point for each attribute according to information gain, selects the best attribute from multiple attributes as the root node, divides the left and right subtrees according to the best classification division point corresponding to the root node attribute, iterates and recursively to form a random forest, until all attribute features of the related videos are of the same category or the related video samples are empty, and outputs the category as the popularity of the to-be-predicted video.

[0044] Specifically, in order to determine the selection of the root node, we need to introduce the concepts of information entropy and information gain, denoted by p i (i=1,2,…,ξ) is the proportion of the i-th sample in the video sample set Ω to be uploaded, ξ represents the total number of sample types in Ω. According to the definition of video popularity, there are three types of samples (low, medium, high), so ξ=3. Then the information entropy of Ω is defined as:

[0045]

[0046]

[0047] The above |Ω| is the number of samples in the video set Ω to be uploaded. is the sample set Ω belonging to class i The number of samples, class i ∈{low,medium,high}, it is stipulated that if p i = 0, then p i log2pi = 0. Note Ave feature Represents the mean of relevant video features of the uploaded video, where feature∈{VC,LC,FO,CM,CI,FC,SC}. Including the number of followers of the uploader (Followers Count, FO), the number of views (View Count, VC), the number of comments (Comment Count, CM), the number of likes (Like Count, LC), the number of favorites (Favorite Count, FC), the number of shares (Share Count, SC), and the number of coins (CoinsCount, CI). For example, given the sample attribute Ave VC , the other sample attributes are the same, assuming Ave VC There are J different values ​​in Ω. Sort these values ​​from small to large and record them as Based on the partition point j, Ω can be divided into subsets and in Indicates that in the attribute Ave VC The subsample set is no larger than j. Indicates that in the attribute Ave VC The subset sample set greater than j, And find the best partition point through information gain.

[0048] The method of finding the best classification point for each attribute according to information gain includes maximizing the information gain and calculating the corresponding optimal classification point, which is expressed as:

[0049]

[0050]

[0051] Where, Gain(Ω,Ave feature ,j) represents the sample set Ω in attribute Ave feature Based on the information gain of the partition point j, For the sample set Ω under attribute Ave feature The optimal partition point on ; Indicates the optimal classification point corresponding to the attribute feature, From the above formula, we can get: When there are multiple solutions, one of them is randomly selected. Ω represents the sample set of related videos of the video to be predicted, Ave feature represents the sample of attribute feature; j represents the ranking of attribute feature; Gain represents the information gain function; Ent(Ω) represents the information entropy of sample set Ω; Indicates that in the attribute Ave featureThe subsample set is no larger than j. Indicates that in the attribute Ave feature The subsample set whose value is greater than j; |Ω| represents the number of samples of the related video of the video to be predicted.

[0052] The optimal classification point corresponding to the root node attribute includes maximizing the information gain corresponding to the optimal classification point and calculating the corresponding optimal classification, which is expressed as:

[0053]

[0054]

[0055] Among them, the meaning of information gain is the difference in information entropy before and after the sample set is divided according to the attribute value. The larger the difference, the smaller the information entropy of the subset after division, that is, the higher the purity of the subset. Therefore, the above process is the optimal division when a single attribute is given. represents the optimal classification point corresponding to the attribute feature, Ω represents the sample set of related videos of the video to be predicted, and Ave feature represents the sample of attribute feature; j represents the ranking of attribute feature; Gain represents the information gain function; Ent(Ω) represents the information entropy of sample set Ω; Indicates that in the attribute Ave feature No more than The sub-sample set of Indicates that in the attribute Ave feature Greater than subsample set; |Ω| represents the number of samples of the related video of the video to be predicted.

[0056] After determining the root node attribute, the left subtree and right subtree of the decision tree use the root node attribute corresponding to and Determine the secondary node, and repeat this process recursively until all sample attributes in the sub-sample set are of the same class or the sub-sample set is empty. In actual construction, it is also necessary to control the depth of the tree to prevent overfitting. If the subset belongs to more than one class, the class with the largest number of samples is used as the final result, and the splitting stops.

[0057] Specifically, the convergence conditions in the decision tree construction process include until all attribute features of the video are of the same category or the subset video samples are empty; it also includes achieving convergence by controlling the depth of the decision tree. If the tree depth reaches a set threshold, and the subset at this time belongs to a collection of non-single-class samples, that is, the subset contains samples of different popularity, then the class with the largest number of popular samples in the sample set is taken as the final result, and the splitting is stopped.

[0058] The following are some results of the implementation method of the present invention, and the embodiment conditions include K=15, Z=5, full features, and a tree depth of 10. In Table 1, the present invention implements three algorithms, namely decision tree, random forest, and multi-layer perceptron, and obtains accuracy rates of 80.18%, 85.97%, and 80.30%, respectively. Among them, the random forest has the highest accuracy rate, which is because the errors caused by multiple single decision trees are compensated during the construction of the random forest.

[0059] Table 1 Comparison of prediction results of the method of the embodiment of the present invention

[0060] algorithm Accuracy Decision Tree (J48) 80.18％ Random Forest (RF) 85.97％ Multilayer Perceptron (MLP) 80.30％

[0061] The embodiment method of the present invention explores the influence of different Z values ​​on the prediction results, and the results are shown in Table 2. In the embodiment method, K=15 is set, all features are used, and the algorithm is a decision tree. It can be seen from Table 2 that the influence of Z value on the result performance fluctuates within 0-1%.

[0062] Table 2 Analysis results of the method of the embodiment of the present invention for the selection of Z (K = 15) value

[0063] Z 1 3 5 7 9 Accuracy 79.12％ 79.34％ 80.18％ 80.07％ 80.04％

[0064] The embodiment method of the present invention explores the influence of different K values ​​on the prediction results, and the results are shown in Table 3. In the embodiment method, Z=5 is set, all features are used, and the algorithm is a decision tree. It can be seen from Table 3 that the K value has a greater impact on the results than the Z value, and the fluctuation range is between 0 and 5.06%. The appropriate range of K value is [9,19].

[0065] Table 3 Analysis results of the method of the embodiment of the present invention for the selection of K (Z = 5) value

[0066] K 1 3 5 7 9 11 13 15 17 19 Accuracy 65.12％ 67.34％ 74.18％ 77.07％ 80.04％ 81.78％ 80.50％ 80.18％ 80.27％ 80.1％

[0067] In the description of the present invention, it is necessary to understand that the terms "coaxial", "bottom", "one end", "top", "middle", "the other end", "upper", "one side", "top", "inside", "outside", "front", "center", "both ends" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.

[0068] In the present invention, unless otherwise clearly stipulated and limited, the terms such as "installation", "setting", "connection", "fixation" and "rotation" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral one; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium; it can be the internal connection of two elements or the interaction relationship between two elements. Unless otherwise clearly defined, ordinary technicians in this field can understand the specific meanings of the above terms in the present invention according to the specific circumstances.

[0069] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting video popularity a priori by integrating video text content, characterized in that: include: Obtain video data from the target video website, including text features and popularity numerical features of different attributes, to form a structured data set; Preprocess the video data, extract the video text content, sort the video data in descending order according to the number of views, and mark the popularity classification of the video data; The preprocessed video text content is extracted using the TextRank algorithm, and the extracted keywords are vectorized using the word2vec algorithm; The cumulative calculation method is used for multiple keywords of each video data to calculate the similarity between the word vector of the video to be predicted published by the current author and the word vector of videos published by other authors; Select K other videos with the highest similarity ranking as the first related videos of the video to be predicted, and calculate the mean of the corresponding attribute popularity numerical features of the K related videos respectively; Select the Z videos recently released by the author of the video to be predicted as the second related videos, and calculate the mean of the corresponding attribute popularity numerical features of the Z related videos respectively; The average of the popularity numerical features of the first related video and the second related video corresponding to the attribute of the video to be predicted is calculated again to generate an attribute feature; The attribute features of videos related to the video to be predicted and their corresponding popularity are classified, and the prior popularity of the video to be predicted is predicted using the random forest algorithm based on decision trees.

2. The method for predicting video popularity a priori by integrating video text content according to claim 1, characterized in that: The popularity numerical features of different attributes include the number of followers (Followers Count, FO) of the uploader, the number of views (View Count, VC), the number of comments (Comment Count, CM), the number of likes (Like Count, LC), the number of favorites (Favorite Count, FC), the number of shares (Share Count, SC), and the number of coins (Coins Count, CI); the text features include the video title, video tags, video introduction, and the author's accompanying text description of the video.

3. The method for predicting video popularity a priori based on the fusion of video and text content according to claim 1, characterized in that: The video data is sorted in descending order according to the amount of playback, and the popularity classification of the video data is marked, including classifying the video into high, medium and low according to the amount of playback of the video data, and marking the popularity classification according to high:[c [0.1N] ,c1],medium:[c [0.4N] ,c [0.1N]+1 ],low:[c N ,c [0.4N]+1 ], where c i Indicates the number of views of the video ranked i, i∈[1,2,...,N], [0.1N] indicates the maximum integer value not exceeding 0.1N, [0.4N] indicates the maximum integer value not exceeding 0.4N, [N] indicates the maximum integer value not exceeding N, and N indicates the total number of videos.

4. The method for predicting video popularity a priori by integrating video text content according to claim 1, characterized in that: The method of cumulative calculation is adopted for multiple keywords of each video data to calculate the similarity between the word vector of the video to be predicted published by the current author and the word vector of videos published by other authors, including respectively calculating the similarity between the M keyword word vectors of the video to be predicted and the N keyword word vectors of videos published by other authors, accumulating the similarities of the M keyword word vectors, and calculating the similarity between the word vector of the video to be predicted published by the current author and the word vector of videos published by other authors.

5. The method for predicting video popularity a priori by integrating video text content according to claim 4, characterized in that: The formula used to calculate the similarity of word vectors is expressed as: Among them, key1 represents the keyword of the video to be predicted, vec1 is the keyword word vector corresponding to key1, key2 represents the keyword of the video released by other authors, and vec2 is the keyword word vector corresponding to key2; is the value of the hth dimension of the vec1 vector, It is the value of the h-th dimension of the vec2 vector, where H represents the word vector dimension.

6. The method for predicting video popularity a priori by integrating video text content according to claim 1, characterized in that: The method of using a random forest algorithm based on a decision tree to predict the prior popularity of a video to be predicted includes sorting the attribute features of each attribute of the related videos of the video to be predicted, finding the best classification point for each attribute according to information gain, selecting the best attribute from multiple attributes as the root node, dividing two left and right subtrees according to the best classification point corresponding to the root node attribute, iterating and recursively forming a random forest until all attribute features in the related videos are of the same category or the related video samples are empty, and outputting the category as the popularity of the video to be predicted.

7. The method for predicting video popularity a priori by integrating video text content according to claim 6, characterized in that: The method of finding the best classification point for each attribute according to information gain includes maximizing the information gain and calculating the corresponding optimal classification point, which is expressed as: in, represents the optimal classification point corresponding to the attribute feature, Ω represents the sample set of related videos of the video to be predicted, and Ave feature Represents a sample of attribute feature; j J represents the total number of rankings of attribute features of attribute feature; Gain represents the information gain function; Ent(Ω) represents the information entropy of the sample set Ω; Indicates that in the attribute Ave feature The subsample set is no larger than j. Indicates that in the attribute Ave feature The subsample set whose value is greater than j; |Ω| represents the number of samples of the related video of the video to be predicted.

8. The method for predicting video popularity a priori by integrating video and text content according to claim 6, characterized in that: The optimal classification point corresponding to the root node attribute includes maximizing the information gain corresponding to the optimal classification point and calculating the corresponding optimal classification, which is expressed as: in, represents the optimal classification point corresponding to the attribute feature, Ω represents the sample set of related videos of the video to be predicted, and Ave feature A sample representing an attribute feature; Indicates the order of attribute features. J represents the total number of rankings of attribute features of attribute feature; Gain represents the information gain function; Ent(Ω) represents the information entropy of the sample set Ω; Indicates that in the attribute Ave feature No more than The sub-sample set of Indicates that in the attribute Ave feature Greater than subsample set; |Ω| represents the number of samples of the related video of the video to be predicted.

9. The method for predicting video popularity a priori by integrating video text content according to claim 6, characterized in that: In the process of building a decision tree, the convergence conditions include until all attribute features of the video are of the same category or the subset video samples are empty; it also includes achieving convergence by controlling the depth of the decision tree. If the tree depth reaches the set threshold, and the subset at this time belongs to a collection of non-single-class samples, that is, the subset contains samples of different popularity, then the class with the largest number of popular samples in the sample set is taken as the final result, and the splitting is stopped.

Images