Method for automatically generating short video based on watching behavior data

Abstract

The invention discloses a method for automatically generating a short video based on watching behavior data, and belongs to the technical field of video production. The invention discloses a method for automatically generating a short video based on watching behavior data. The method comprises the following steps: slicing a video into a plurality of video clips with the same length; performing multi-dimensional value analysis on each video clip, wherein the value of each video clip forms a value weight matrix; calculating the value weight of each video clip based on the one-dimensional matrix and the trained clip value model; obtaining watching behavior information of the user; calculating a standardized value score of the video clip based on the viewing behavior information of the user and the value weight matrix; and generating a simplified video according to the standardized value score of the video clip and the video clip. According to the invention, the efficiency of searching and watching the knowledge point video by the user is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of video production, and in particular to a method for automatically generating short videos based on viewing behavior data. Background Art

[0002] With the popularity of online education, the number of teaching video resources collected by schools through live broadcasts and transcriptions has increased. There are a large number of course resources available for students to learn. However, it is difficult for students to locate the course videos with more concise explanations and the knowledge points they are concerned about in the massive video resources. The existing teaching video production system mainly has the following two problems:

[0003] 1) An online course usually lasts one or two hours and contains a lot of content that is irrelevant to the course knowledge points. Students hope to skip content such as roll call, self-study, and discussion and get a refined course video.

[0004] 2) The existing method of extracting knowledge points is basically to manually mark and cut the video after it is collected. Obviously, this method not only consumes a lot of manpower and time, but also has low efficiency and cannot meet the timeliness requirements. The existing course search is to search for information following the course entry, and users cannot quickly retrieve the knowledge point video content. Summary of the invention

[0005] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a method for automatically generating short videos based on viewing behavior data.

[0006] The object of the present invention is achieved through the following technical solution: a method for automatically generating short videos based on viewing behavior data, comprising:

[0007] Slice the video into several video clips of the same length;

[0008] Perform multi-dimensional value analysis on each video clip, and the value of each video clip forms a value weight matrix;

[0009] Calculate the value weight of each video segment based on the one-dimensional matrix and the trained segment value model;

[0010] Obtain user viewing behavior information;

[0011] Calculate the standardized value score of the video clip based on the user's viewing behavior information and the value weight matrix;

[0012] A condensed video is generated based on the normalized value scores of the video segments and the video segments.

[0013] Furthermore, a multi-dimensional value analysis is performed on each video clip, including:

[0014] Extract audio data from video clips;

[0015] Perform noise reduction on audio data;

[0016] Setting one or more volume thresholds;

[0017] Extracting audio features of audio data, the audio features being energy or amplitude;

[0018] An audio analysis score is determined based on audio characteristics and a volume threshold.

[0019] Furthermore, a multi-dimensional value analysis is performed on each video clip, including:

[0020] Detecting the presence of people in a video clip;

[0021] If a person is detected in the video clip, the person weight is set to 1, otherwise the person weight is set to 0. Then the pre-trained behavior recognition model is used to determine whether the person is standing or giving a lecture.

[0022] Calculate the character's behavior matching weight score based on the detection results of the behavior recognition model;

[0023] The video character analysis score of each video clip is calculated based on the character weight and the character's behavior matching weight score. The calculation formula is as follows:

[0024] V i =P i ·W i

[0025] Among them, V i is the video image character analysis score, P i is the character weight, W i A weighted score for the character's behavior matching.

[0026] Furthermore, a multi-dimensional value analysis is performed on each video clip, including:

[0027] Perform speech recognition on the entire video to obtain speech recognition text;

[0028] Determine the video clip corresponding to each sentence based on the timestamp of the speech recognition text;

[0029] The speech recognition result analysis score of the video segment is determined according to whether the video segment has corresponding speech recognition text.

[0030] Furthermore, a multi-dimensional value analysis is performed on each video clip, including:

[0031] Perform speech recognition on the entire video to obtain speech recognition text;

[0032] Segment the speech recognition text according to the timestamp to obtain a number of text segments;

[0033] Determine the corresponding scene based on the content of the text segment;

[0034] Obtaining a first similarity between the text segment and the corresponding scene based on the frequency of occurrence of a preset scene keyword in the text segment;

[0035] Based on the topic similarity, a second similarity between the text segment and the corresponding scene is obtained;

[0036] Get the preset basic weight value of the scene corresponding to the text segment;

[0037] The first similarity, the second similarity and the basic weight value are multiplied to obtain a scene weight value of the text segment;

[0038] Determine the video segment corresponding to the text segment according to the timestamp;

[0039] The scene analysis score of the speech recognition result of the video segment is determined according to the scene weight value of the text segment.

[0040] Furthermore, a multi-dimensional value analysis is performed on each video clip, including:

[0041] Perform speech recognition on the entire video to obtain speech recognition text;

[0042] Extract triple keywords from each sentence in the speech recognition text, and convert the extracted keywords into vector form to obtain keyword vectors;

[0043] Calculate the cosine similarity between the keyword vector and the vector of each knowledge point in the course knowledge graph, and calculate the weight value of the keyword based on the cosine similarity and the basic weight of each knowledge point. The calculation formula is as follows:

[0044]

[0045] Among them, K1 is the weight value of the keyword, I i is the similarity between the keyword and the i-th knowledge point in the knowledge graph, w i is the basic weight of the i-th knowledge point;

[0046] Determine the text content corresponding to each video clip according to the timestamp information of the speech recognition text;

[0047] The knowledge point analysis score of the speech recognition result of the video clip is obtained according to the weight value of the key words in the text content corresponding to the video clip.

[0048] Furthermore, the value weight of each video segment is calculated based on the one-dimensional matrix and the trained segment value model, including:

[0049] Set the fragment value model to a linear regression model, and the parameters of the fragment value model are [a1, a2, a3, a4, a5];

[0050] The calculation formula of value weight W is:

[0051] W=a1·V+a2·P+a3·T+a4·S+a5·K

[0052] Among them, V is the audio analysis score, P is the video character analysis score, T is the speech recognition analysis score, S is the speech recognition result scene analysis score, and K is the speech recognition result knowledge point analysis score;

[0053] Normalize the value weight W.

[0054] Furthermore, the standardized value score of the video clip is calculated based on the user's viewing behavior information and the value weight matrix, including:

[0055] The user viewing hot zone is generated based on the user's reverse dragging behavior and viewing history. The viewing heat value H of each video clip in the user viewing hot zone is i is the number of times the video clip is re-watched by the user or the total re-watching duration. The user viewing hotspot is recorded as a one-dimensional viewing hotspot matrix [H1H2...H n ], where n is the number of video clips;

[0056] A user skip area is generated according to the user's forward dragging behavior, and the ignore value L of each video segment in the user skip area is i is the number of times the video segment is skipped by the user or the total skipping duration. The user skipped area is recorded as a one-dimensional viewing ignored interval matrix [L1L2...L n ], where n is the number of video clips;

[0057] According to the value weight matrix of the entire video [W1W2...W n ], watch the hot zone matrix [H1H2...H n ] and viewing the ignored interval matrix [L1L2...L n ] to perform weighted calculation to generate the final value score of each video clip. The calculation formula for weighted calculation is:

[0058] V i =W i +α·Hi-β·Li

[0059] Among them, V iis the final value score of the i-th video clip; α and β are weight coefficients;

[0060] The calculated V i The values ​​are normalized to the range of [0,1] to obtain a normalized value score for each video clip.

[0061] Further, a condensed video is generated according to the standardized value scores of the video segments and the video segments, including:

[0062] Determine the video clips whose normalized value scores are greater than or equal to a preset threshold as high-value clips;

[0063] Determine a video segment whose normalized value score is less than a preset threshold as a low-value segment;

[0064] generating a condensed video based on the video segments determined to be high-value segments;

[0065] Establish the relationship index between the streamlined video and the original video.

[0066] The beneficial effects of the present invention are as follows: the present invention provides a method for automatically generating short videos based on viewing behavior data, establishes a special context data model according to the usual course teaching structure, combines the user viewing behavior data, combines the course intelligent processing result data, intercepts the course video clips, and automatically generates new course video clips. This method improves the defects of manually marking and cutting knowledge point shots in the prior art, and can automatically edit short videos in the context of a large number of recorded class videos in schools, saving a lot of manpower. The method implements recommendations based on user behavior data analysis, establishes a video content relationship index through a content relationship map, and generates a collection of exquisite course videos, which improves the efficiency of users retrieving knowledge point videos and watching knowledge point videos. The streamlined video can also increase the playback speed when the user is watching, and by editing out the content of invalid video parts, a lot of storage space can also be saved. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 A flow chart of the method of the present invention. DETAILED DESCRIPTION

[0068] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0069] See also Figure 1 , this embodiment provides a method for automatically generating short videos based on viewing behavior data:

[0070] like Figure 1 As shown, a method for automatically generating short videos based on viewing behavior data includes steps S100 to S600.

[0071] Step S100: Slice the video into a number of video segments of the same length.

[0072] For example, the video is logically segmented into content segments every 5 seconds, and the video is serialized into a one-dimensional array based on the time dimension. The number of items in the one-dimensional array is equal to the number of video segments generated by the video.

[0073] Step S200: Perform multi-dimensional value analysis on each video clip respectively, and the value of each video clip constitutes a value weight matrix.

[0074] In one embodiment, a multi-dimensional value analysis is performed on each video clip, including: audio analysis, video image character analysis, speech recognition result analysis, speech recognition result scene analysis, and speech recognition result knowledge point analysis.

[0075] Let the audio analysis score be V, the video character analysis score be P, the speech recognition analysis score be T, the speech recognition result scene analysis score be S, and the speech recognition result knowledge point analysis score be K. The value analysis results of each video clip constitute a one-dimensional value analysis matrix of [V, P, T, S, K].

[0076] In one embodiment, audio analysis of a video clip includes: extracting audio data from the video clip; performing noise reduction processing on the audio data, for example, using Librosa or other audio processing libraries to perform noise reduction processing on the audio data, such as using spectral subtraction or wavelet transform to perform noise reduction on the audio data, the denoised audio data can more clearly reflect the actual content and improve the accuracy of subsequent analysis; setting one or more volume thresholds, such as silence (0dB), low voice (20dB), normal speech (60dB), and loud voice (100dB); extracting audio features of the audio data, the audio features are energy or amplitude, for example, using Librosa to extract the energy or amplitude of the audio data; determining the audio volume level according to the audio features and the volume threshold, that is, obtaining the audio analysis score of the video clip, for example, if the energy of the audio is greater than 40dB, it is considered to be normal speech and is assigned a value of 0.8; if the energy is less than 20dB, it is considered to be silence and is assigned a value of 0. Thus, a valid value in the interval [0,1] is obtained according to the clarity and analyzability of the sound.

[0077] In one embodiment, performing a video character analysis on a video clip includes: detecting whether there is a person in the video clip, for example, using an object detection algorithm in OpenCV or a deep learning framework (such as TensorFlow, PyTorch) to detect each person in the video; if a person is detected in the video clip, setting a person weight P i =1, otherwise set the character weight P i = 0, and then use the pre-trained behavior recognition model to determine whether the person is standing and whether he is giving a lecture. The behavior recognition model can be based on posture estimation (such as OpenPose) or action recognition algorithm; calculate the behavior matching weight score (confidence) of the person according to the detection result of the behavior recognition model. For example, if the behavior recognition model predicts that a person is giving a lecture and the confidence is 0.9, then the behavior matching weight score W i =0.9, combined with P i and W i , calculate the video screen character analysis score of each video clip, for example, the video screen character analysis score V of each video clip i It can be defined as V i =P i ·W i .

[0078] In one embodiment, performing speech recognition result analysis on a video clip includes: performing speech recognition on the entire video to obtain speech recognition text, for example, extracting audio data from the entire video and converting the extracted audio data into text; determining the video clip corresponding to each sentence according to the timestamp of the speech recognition text; and determining the speech recognition result analysis score of the video clip according to whether the video clip has a corresponding speech recognition text, for example, if the video clip has a corresponding speech recognition text, its speech recognition result analysis score is 1, otherwise its speech recognition result analysis score is 0.

[0079] In one embodiment, the scene analysis of speech recognition results of the video clip includes: performing speech recognition on the entire video to obtain speech recognition text, for example, extracting audio data from the entire video, and converting the extracted audio data into text. The speech recognition text is segmented according to the timestamp to obtain a number of text segments, for example, the speech recognition text with continuous timestamps is divided into a text segment, and a single text segment is no longer than 10 sentences. The corresponding scene is determined according to the content of the text segment. For example, when the speech text mentions semantic scenes such as "mentioned in the last class" and "let's review", this content is recorded as a class review. According to the defined course results, it can be divided into scenes such as class review, teaching new knowledge, questioning exercises, classroom interaction, pre-class preparation, roll call, key explanation, etc. A set of scene keywords or scene phrases is set for each predefined scene, and then the frequency of occurrence of these scene keywords in the text segment is counted, and the first similarity between the text segment and the corresponding scene is obtained according to the frequency of occurrence of the scene keywords, wherein the more the scene keywords appear, the greater the score of the first similarity (i.e., the higher the similarity). Through topic modeling techniques such as LDA (Latent Dirichlet Allocation), the main topic of the text segment is found, and the similarity is matched with the topic of the predefined scene to obtain the second similarity between the text segment and the corresponding scene. A high score of the second similarity means that the text segment is very compatible with a specific scene, but this is only part of the factor that determines the final weight value. A basic weight value is set for each predefined scene. For example, a class review may be assigned a lower basic weight value because this part of the content is relatively basic; while a key explanation may be assigned a higher basic weight value because the knowledge points it contains are more important. For each text segment, the first similarity, the second similarity and the basic weight value of the corresponding scene are multiplied to obtain the scene weight value of the text segment. The video segment corresponding to the text segment is determined according to the timestamp. The scene analysis score of the speech recognition result of the video segment is determined according to the scene weight value of the text segment. For example, the scene analysis score of the speech recognition result of the video segment is calculated by the average, maximum or other aggregation functions of the scene weight values ​​of the corresponding text segment.

[0080] In one embodiment, the knowledge point analysis of the speech recognition results of the video clip includes: performing speech recognition on the entire video to obtain speech recognition text, such as extracting audio data from the entire video, and converting the extracted audio data into text. Extract triple keywords from each sentence in the speech recognition text respectively; then use a pre-trained language model (such as BERT, Word2Vec, etc.) to convert these keywords into vector form to obtain a keyword vector. Calculate the cosine similarity I between the keyword vector and the knowledge point vectors in the course knowledge graph. The weight value K1 of the keyword can be calculated by weighted average, that is: Among them, Ii is the similarity between the keyword and the i-th knowledge point in the knowledge graph, w i is the basic weight of the i-th knowledge point, the basic weight w i It can be pre-set according to the importance of the knowledge point in the course. The calculated weight value K1 of each keyword is mapped back to the original speech recognition text, that is, the keywords extracted from each sentence will have a corresponding weight value. According to the timestamp information of the speech recognition text, the text content corresponding to each video clip is determined. The speech recognition result knowledge point analysis score of the video clip is obtained according to the weight value of the keyword in the text content corresponding to the video clip. For example, the speech recognition result knowledge point analysis score of the video clip can be calculated according to the average, maximum or other aggregation functions of all keyword weight values ​​in the video clip, such as using the average method: Among them, K1 j is the weight value of the jth keyword in the video clip, and m is the number of keywords in the video clip.

[0081] Step S300: Calculate the value weight of each video segment based on the one-dimensional matrix and the trained segment value model.

[0082] For example, supervised AI expert model training is performed on the value matrix of each video clip. By setting different coefficients on each weight value, the value weight of a video clip can be calculated, recorded as W, and the entire video can be recorded as a one-dimensional value weight matrix [W1, W2, ... W n ], the number of entries in the value weight matrix is ​​equal to the number of video segments generated by video segmentation.

[0083] Specifically, assuming that the fragment value model is a linear regression model and the parameters of the fragment value model are [a1, a2, a3, a4, a5], the value weight W can be calculated by the following formula:

[0084] W=a1·V+a2·P+a3·T+a4·S+a5·K

[0085] In order to ensure that the value weight W is within a reasonable range (for example, [0, 1]), the calculation result can be normalized. Common normalization methods include minimum-maximum normalization, and the calculation formula is as follows:

[0086]

[0087] Repeat the above calculation process for each video clip, and finally generate the value weight matrix [W1W2...W n ], where n is the number of video clips.

[0088] Step S400: Obtain the user's viewing behavior information.

[0089] When a user watches a video, after the user enters the playback page, the user's behavior of clicking on the video to learn is recorded through xAPI. When the user watches a video, the user's viewing time is recorded regularly (such as every 10 seconds), the total viewing time of the user is counted, and the user's pause, play, forward drag, reverse drag, etc. are recorded. The start and end points of the forward drag video are recorded, and the time interval of the forward drag is recorded as the area of ​​the video that the user wants to ignore. The start and end points of the reverse drag are recorded, and the time interval of the reverse drag is marked as the area that the user wants to watch repeatedly.

[0090] Step S500: Calculate the standardized value score of the video clip based on the user's viewing behavior information and the value weight matrix.

[0091] According to the user's reverse dragging behavior (indicating that the user has watched repeatedly) and the user's viewing history, the video interval watched by the user is superimposed according to the video clips of the video to generate a user viewing hot zone. The viewing heat value H of each video clip in the user viewing hot zone is i is the number of times the video clip is re-watched by the user or the total re-watching duration. The user viewing hotspot is recorded as a one-dimensional viewing hotspot matrix [H1H2...H n ], where n is the number of video clips.

[0092] A user skip area is generated according to the user's forward dragging behavior (skipping viewing), and the ignore value L of each video segment in the user skip area is i is the number of times the video segment is skipped by the user or the total skipping duration. The user skipped area is recorded as a one-dimensional viewing ignored interval matrix [L1L2...L n ], where n is the number of video clips.

[0093] According to the value weight matrix of the entire video [W1W2...W n ], watch the hot zone matrix [H1H2...H n ] and viewing the ignored interval matrix [L1L2...L n ] to perform weighted calculation to generate the final value score of each video clip. The calculation formula for weighted calculation is:

[0094] V i =W i +α·Hi-β·Li

[0095] Among them, V i is the final value score of the i-th video clip; α and β are weight coefficients used to adjust the influence of the viewing hot zone and the viewing ignored interval.

[0096] The calculated V iThe value is normalized to the range of [0,1] to obtain the standardized value score V for each video clip. i ′, for subsequent processing. Use the minimum-maximum normalization method: Among them, V max and V min are the maximum and minimum values ​​of the final value scores of all fragments, respectively.

[0097] Step S600: Generate a condensed video according to the standardized value score of the video segment and the video segment.

[0098] A threshold T is set to distinguish high-value segments from low-value segments. Generally, the threshold T is the average value of the standardized value scores of all video segments, that is, T = mean (V'1, V'2, ..., V' n ). If the standardized value score V of a video clip i ′ is greater than or equal to the threshold T, then the video clip is a high-value clip; if the standardized value score V i ′ is less than the threshold T, then the video segment is a low-value segment; a simplified video is generated based on the video segment determined to be a high-value segment.

[0099] For example, the video segments determined to be high-value segments are marked as 1, and the video segments determined to be low-value segments are marked as 0, and the recommended region matrix [R1R2...R n ], where R i Indicates whether the i-th video segment is recommended. According to the recommended region matrix [R1R2...R n ], extract all video clips marked as 1. Merge continuous high-value clips into a clipping region. Use video editing tools (such as FFmpeg, etc.) to extract corresponding video clips according to the clipping region, splice the extracted video clips together to generate a new streamlined video. Generate a relationship index between the streamlined video and the original video, record the time position of each streamlined video in the original video, and users can quickly locate the corresponding part of the original video when watching the streamlined video.

[0100] The above is only a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, and should not be regarded as excluding other embodiments, but can be used in various other combinations, modifications and environments, and can be modified within the scope of the concept described herein through the above teachings or the technology or knowledge of the relevant field. The changes and modifications made by those skilled in the art shall not deviate from the spirit and scope of the present invention, and shall be within the scope of protection of the claims attached to the present invention.

Claims

1. A method for automatically generating short videos based on viewing behavior data, characterized in that: include: Slice the video into several video clips of the same length; Perform multi-dimensional value analysis on each video clip, and the value of each video clip forms a value weight matrix; Calculate the value weight of each video segment based on the one-dimensional matrix and the trained segment value model; Obtain user viewing behavior information; Calculate the standardized value score of the video clip based on the user's viewing behavior information and the value weight matrix; A condensed video is generated based on the normalized value scores of the video segments and the video segments.

2. A method for automatically generating short videos based on viewing behavior data according to claim 1, characterized in that: Conduct multi-dimensional value analysis on each video clip, including: Extract audio data from video clips; Perform noise reduction on audio data; Setting one or more volume thresholds; Extracting audio features of audio data, the audio features being energy or amplitude; An audio analysis score is determined based on audio characteristics and a volume threshold.

3. The method for automatically generating short videos based on viewing behavior data according to claim 1, characterized in that: Conduct multi-dimensional value analysis on each video clip, including: Detecting the presence of people in a video clip; If a person is detected in the video clip, the person weight is set to 1, otherwise the person weight is set to 0. Then the pre-trained behavior recognition model is used to determine whether the person is standing or giving a lecture. Calculate the character's behavior matching weight score based on the detection results of the behavior recognition model; The video character analysis score of each video clip is calculated based on the character weight and the character's behavior matching weight score. The calculation formula is as follows: V i =P i ·W i Among them, V i is the video character analysis score, P i is the character weight, W i The weighted score for the character's behavior matching.

4. The method for automatically generating short videos based on viewing behavior data according to claim 1, characterized in that: Conduct multi-dimensional value analysis on each video clip, including: Perform speech recognition on the entire video to obtain speech recognition text; Determine the video clip corresponding to each sentence based on the timestamp of the speech recognition text; The speech recognition result analysis score of the video segment is determined according to whether the video segment has corresponding speech recognition text.

5. The method for automatically generating short videos based on viewing behavior data according to claim 1, characterized in that: Conduct multi-dimensional value analysis on each video clip, including: Perform speech recognition on the entire video to obtain speech recognition text; Segment the speech recognition text according to the timestamp to obtain a number of text segments; Determine the corresponding scene based on the content of the text segment; Obtaining a first similarity between the text segment and the corresponding scene based on the frequency of occurrence of a preset scene keyword in the text segment; Based on the topic similarity, a second similarity between the text segment and the corresponding scene is obtained; Get the preset basic weight value of the scene corresponding to the text segment; The first similarity, the second similarity and the basic weight value are multiplied to obtain a scene weight value of the text segment; Determine the video segment corresponding to the text segment according to the timestamp; The scene analysis score of the speech recognition result of the video segment is determined according to the scene weight value of the text segment.

6. The method for automatically generating short videos based on viewing behavior data according to claim 1, characterized in that: Conduct multi-dimensional value analysis on each video clip, including: Perform speech recognition on the entire video to obtain speech recognition text; Extract triple keywords from each sentence in the speech recognition text, and convert the extracted keywords into vector form to obtain keyword vectors; Calculate the cosine similarity between the keyword vector and the vector of each knowledge point in the course knowledge graph, and calculate the weight value of the keyword based on the cosine similarity and the basic weight of each knowledge point. The calculation formula is as follows: Among them, K1 is the weight value of the keyword, I i is the similarity between the keyword and the i-th knowledge point in the knowledge graph, w i is the basic weight of the i-th knowledge point; Determine the text content corresponding to each video clip according to the timestamp information of the speech recognition text; The knowledge point analysis score of the speech recognition result of the video clip is obtained according to the weight value of the key words in the text content corresponding to the video clip.

7. The method for automatically generating short videos based on viewing behavior data according to claim 1, characterized in that: The value weight of each video segment is calculated based on the one-dimensional matrix and the trained segment value model, including: Set the fragment value model to a linear regression model, and the parameters of the fragment value model are [a1, a2, a3, a4, a5]; The calculation formula of value weight W is: W=a1·V+a2·P+a3·T+a4·S+a5·K Among them, V is the audio analysis score, P is the video character analysis score, T is the speech recognition analysis score, S is the speech recognition result scene analysis score, and K is the speech recognition result knowledge point analysis score; Normalize the value weight W.

8. The method for automatically generating short videos based on viewing behavior data according to claim 1, characterized in that: The standardized value score of the video clip is calculated based on the user's viewing behavior information and the value weight matrix, including: The user viewing hot zone is generated based on the user's reverse dragging behavior and viewing history. The viewing heat value H of each video clip in the user viewing hot zone is i is the number of times the video clip is re-watched by the user or the total re-watching duration. The user viewing hotspot is recorded as a one-dimensional viewing hotspot matrix [H1 H2...H n ], where n is the number of video clips; A user skip area is generated according to the user's forward dragging behavior, and the ignore value L of each video segment in the user skip area is i is the number of times the video segment is skipped by the user or the total skipping duration. The user skipped area is recorded as a one-dimensional viewing ignored interval matrix [L1 L2...L n ], where n is the number of video clips; According to the value weight matrix of the entire video [W1 W2...W n ], watch the hot zone matrix [H1 H2...H n ] and viewing the ignored interval matrix [L1 L2 ... L n ] to perform weighted calculation to generate the final value score of each video clip. The calculation formula for weighted calculation is: V i =W i +α·Hi-β·Li Among them, V i is the final value score of the i-th video clip; α and β are weight coefficients; The calculated V i The values ​​are normalized to the range of [0,1] to obtain a normalized value score for each video clip.

9. The method for automatically generating short videos based on viewing behavior data according to claim 1, characterized in that: Generate a condensed video based on the normalized value scores of the video clips and the video clips, including: Determine the video clips whose normalized value scores are greater than or equal to a preset threshold as high-value clips; Determine a video segment whose normalized value score is less than a preset threshold as a low-value segment; generating a condensed video based on the video segments determined to be high-value segments; Establish the relationship index between the streamlined video and the original video.

Images