A video ringback tone playing method based on off-hook duration prediction algorithm
Through the off-hook time prediction algorithm and video exciting clip extraction technology, the problem of incomplete video ringtone playback is solved, data quality and prediction accuracy are improved, and users can receive rich video information.
Patent Information
- Application Number
- CN202210305947.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-25
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2042-03-25
AI Technical Summary
The existing technology lacks effective prediction of the duration of being called off-hook, resulting in the incomplete playback of the video ringtone, the received information is incomplete, and the data quality problem cannot be effectively solved.
The video ringtone playback method based on the off-hook time prediction algorithm is adopted, and the user's off-hook time is predicted through data cleaning and machine learning regression models, and combined with the exciting video clip extraction technology, new video ringtone content is generated.
It has achieved the improvement of the integrity and richness of the video ringtone content, solved the data quality problems and improved prediction accuracy, and ensured that the caller received more key video information.
Smart Images

Figure CN114928675B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of video color ring back tone extraction, in particular to a video color ring back tone playing method based on an off-hook duration prediction algorithm. Background Art
[0002] Currently, telecommunications calls have the following shortcomings:
[0003] 1. Lack of prediction of the off-hook duration of the called party (in the past, the off-hook duration of the corresponding user was directly taken as the duration of a certain connection or the average of multiple calls);
[0004] 2. Solve data quality problems such as missing values, outliers, and no data based on traditional data cleaning algorithms;
[0005] 3. The selection of features is crucial and has a huge impact on the prediction results;
[0006] 4. The calling user cannot play the entire video ringback tone completely, cannot receive the key segment of the video ringback tone, and the received information is incomplete. Summary of the invention
[0007] In order to address the deficiencies of the prior art, the present invention proposes a video ringback tone playback method based on an off-hook duration prediction algorithm, which adopts video highlight segment extraction technology to compile new video ringback tone content according to the predicted off-hook duration of the called user, so that the user can view and receive more key video content information.
[0008] The video ringback tone playing method based on off-hook duration prediction algorithm of the present invention comprises the following steps:
[0009] S1, data acquisition step, obtains all necessary data feature information from the user data background and outputs it in CSV data format;
[0010] S2. Complete quality control on the acquired user feature data, solve the problems of missing values, abnormal values, and no data in the user data information through traditional data cleaning algorithms, and obtain user data information that has passed quality control;
[0011] S3, performing key feature screening on the acquired quality control data, and extracting corresponding key feature data as regression fitting data;
[0012] S4, calculation of the called user off-hook duration prediction algorithm;
[0013] S5. Extract and play the highlight segments of the video ring back tone according to the predicted off-hook time of the called user.
[0014] Preferably, the specific steps of S1 are as follows:
[0015] Ⅰ. First, open the database query window and input the Structured Query Language statement of the data to be exported;
[0016] Ⅱ. Copy all the queried data in the search results and save the results as Comma-SeparatedValues CSV text file and export it.
[0017] Preferably, the method for S2 to obtain the quality-controlled user data information is as follows:
[0018] Ⅰ. Use neighbor interpolation and average interpolation to solve missing values. Neighbor interpolation is to use adjacent data values for interpolation, and average interpolation is to use the average value of the dimension feature for interpolation;
[0019] Ⅱ. Use the self-designed threshold to judge the abnormal value. The maximum length of the phone call is 54 seconds. If it exceeds the threshold, it is an abnormal value. At the same time, negative values and abnormal values are also abnormal values.
[0020] III. Fill in the call connection time of users without data by using the mean or median of the connection time of all users.
[0021] Preferably, the key feature data in S3 include: 1. Off-hook time; 2. Time point of the day; 3. Gender; 4. Monday to weekend; 5. True value of the user's call connection time.
[0022] Preferably, the steps of the called user off-hook duration prediction algorithm in S4 are as follows:
[0023] Ⅰ. Through years of experience in communications and observation and analysis of big data, we screened the key feature information of users' off-hook duration and established a multi-dimensional feature data set for off-hook duration prediction. At the same time, some semantic features need to be converted into corresponding digital features.
[0024] II. Label the established user multi-dimensional feature data set with corresponding labels, so that each feature vector has a corresponding true value label of off-hook duration. The data set contains features of four dimensions and corresponds to the true value of a call connection duration.
[0025] III. The data set is used as the input of the Lasso Regression model. After model training and parameter updating, the final model parameters are obtained, and the regression prediction results are obtained. The regression results are the prediction results of the user's call connection time. The regression model is a trained parameter model obtained from multidimensional feature data. The regression error formula of Lasso Regression is as follows:
[0026]
[0027] IV. Send the test set to the trained model for prediction and judge the model regression effect, mainly through the mean square error MeanSquared Error, MSE, root mean squared error root mean squared error, RMSE, or R-square coefficient for evaluation. The formula is as follows:
[0028]
[0029] Preferably, the specific steps of S5 are as follows:
[0030] Ⅰ. By splitting the short video into frames, using the dimensionality reduction algorithm to reduce the data dimension of each frame and then putting it into an array, a two-dimensional feature vector representing the entire video data is obtained, and a dimension representing time is added to the feature vector according to the timing information of each frame to obtain a three-dimensional feature vector. This three-dimensional feature vector represents the characteristic information of the short video;
[0031] Ⅱ. Based on the k-means algorithm, the three-dimensional feature vector is clustered unsupervisedly based on the time characteristics while retaining the time series information. The number of clustering categories is also adaptive, so as to extract multiple video clips, which are the highlights of the short video.
[0032] III. Based on the predicted duration of the user's call, the video clips are spliced in chronological order, and a wonderful video of not less than and the most recent predicted duration is integrated according to the duration of each video clip. By compressing and extracting the video ringtone content, the user on the call side can obtain as much video information as possible in a limited time.
[0033] Preferably, the k-means algorithm steps are as follows:
[0034] Ⅰ. Evenly select n cluster centers in the time series, where n is one third of the total number of frames;
[0035] Ⅱ. Calculate the average value in the neighborhood of each cluster center point t, where t is half of the distance between the centers;
[0036] III. If the absolute value of the difference between the average values of adjacent clusters is less than the set threshold T, the two clusters are merged;
[0037] IV. Iteratively repeat steps II and III until no new clusters are generated;
[0038] V. According to the number of sample points in each cluster, take the top k clusters as the highlights of the short video.
[0039] The beneficial effects of the present invention are as follows:
[0040] The present invention utilizes traditional data cleaning technology to solve the problem of video ringback tone user data quality caused by data collection errors and user information defects, and obtains good user data information; in order to solve the problem of predicting the user off-hook time, key multidimensional features are selected and combined with a machine learning regression model to complete the fitting prediction; based on the estimation of the user off-hook time, the problem of estimating the video ringback tone push time is solved, and in the ringback tone push process, new video ringback tone content can be generated according to the video content semantics, and meaningful video content information is pushed to the caller.
[0041] For the first time, the prediction of connection time was completed based on data cleaning and machine learning regression algorithm; four key features were selected through big data analysis, which played an important role in improving prediction accuracy; Lasso regression has strong sparsity, and this algorithm is used to solve the processing of feature coefficients with strong correlation; the k-means algorithm is used to solve the extraction of video highlights, and the algorithm has a fast calculation speed and good actual extraction effect. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] The present invention will be further described below in conjunction with the accompanying drawings.
[0043] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION
[0044] In order to make the purpose, technical means and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are 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.
[0045] The embodiment of the present invention provides a video ringback tone playing method based on an off-hook duration prediction algorithm, adopts a video highlight segment extraction technology, and compiles new video ringback tone content according to the off-hook duration prediction of the called user, so that the user can view and receive more key video content information.
[0046] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.
[0047] like Figure 1 As shown, the video ringback tone playing method based on off-hook duration prediction algorithm of the present invention comprises the following steps:
[0048] S1. Data acquisition step: obtain all necessary data feature information from the user data background and output it in CSV data format. Specific steps: first open the database query window, output the Structured Query Language (SQL) statement that needs to export data, then copy all the queried data in the search results and save the results as a Comma-Separated Values (CSV) text file and export it.
[0049] (2) Complete quality control on the acquired user feature data, and solve the problems of missing values, abnormal values, and no data in the user data information through traditional data cleaning algorithms. The steps to obtain user data information that has passed quality control are as follows:
[0050] 1. Use neighbor interpolation and average interpolation to solve missing values. Neighbor interpolation is to use adjacent data values for interpolation, and average interpolation is to use the average value of the dimension feature for interpolation;
[0051] 2. Use the self-designed threshold to determine the outliers. The maximum length of a phone call is 54 seconds. If it exceeds the threshold, it is an outlier. Negative values and abnormal values are also outliers.
[0052] 3. The call connection time of users without data is filled in by using the mean or median of the connection time of all users.
[0053] S3. Perform key feature screening on the acquired quality control data and extract the corresponding key feature data as regression fitting data. The key feature data include: 1. Off-hook time; 2. Time of day (6-12 am, 12-19 pm, evening & 19-6 am); 3. Gender; 4. Monday to weekend (weekdays or weekends); 5. The true value of the user's call connection time.
[0054] S4. The specific steps of the called user off-hook duration prediction algorithm are as follows:
[0055] First, through our many years of experience in communications and observation and analysis of big data, we screened the key feature information of the user's off-hook duration and established a multi-dimensional feature dataset for off-hook duration prediction. At the same time, some semantic features need to be converted into corresponding digital features. See the table below.
[0056] User data quality control table
[0057] user X1 (call) X2(Gender) X3(time point) X4 (Is it a working day) Y(On) A0 50s 1 1 1 5s A1 30s 2 3 1 10s A2 120s 1 2 2 20s
[0058] Then, the established user multi-dimensional feature data set is labeled with corresponding labels, so that each feature vector has a corresponding true value label of the off-hook duration. The data set contains four dimensional features and corresponds to the true value of the call connection duration.
[0059] Next, the data set is used as the input of the Lasso Regression model. After model training and parameter update, the final model parameters are obtained, and the regression prediction results are obtained. The regression results are the prediction results of the user's call connection time. The regression model is a trained parameter model obtained from multi-dimensional feature data. The regression error formula of Lasso Regression is as follows:
[0060]
[0061] Finally, the test set is sent to the trained model for prediction. The regression effect of the model is mainly evaluated by the mean square error (MSE), root mean square error (RMSE), or R-square coefficient. The formula is as follows:
[0062]
[0063] S5. Extracting highlights from the video ring back tone according to the predicted off-hook duration of the called user. The specific steps are as follows:
[0064] First, the present invention decomposes the short video into frames, uses a dimensionality reduction algorithm to reduce the data dimension of each frame, and then puts it into an array to obtain a two-dimensional feature vector representing the entire video data. And according to the timing information of each frame, a dimension representing time is added to the feature vector to obtain a three-dimensional feature vector. This three-dimensional feature vector represents the feature information of the short video.
[0065] Secondly, the algorithm is improved based on the k-means algorithm for short video features. The traditional k-means algorithm performs iterative clustering based on the distance from the sample point to the center of the cluster, but this method will disrupt its timing information when applied to short video data. The improved clustering algorithm can perform unsupervised clustering of the three-dimensional feature vector based on time characteristics while retaining the timing information. The number of clustering categories is also adaptive, thereby extracting multiple video clips, which are the highlights of the short video.
[0066] The specific steps of the improved k-means algorithm are as follows:
[0067] 1. Evenly select n cluster center points in the time series, where n is one third of the total number of frames;
[0068] 2. Calculate the average value in the neighborhood of each cluster center point t, where t is half the distance between the centers;
[0069] 3. If the absolute value of the difference between the average values of adjacent clusters is less than the set threshold T, the two clusters are merged;
[0070] 4. Repeat steps 2 and 3 iteratively until no new clusters are generated;
[0071] 5. According to the number of sample points in each cluster, take the top k clusters as the highlights of the short video.
[0072] Finally, based on the predicted duration t of the user's call, the video clips are spliced in time sequence, and a wonderful video with a duration of not less than and closest to the predicted duration t is integrated according to the duration of each video clip. By compressing and extracting the video ringback tone content, the user on the call side can obtain as much video information as possible in a limited time.
[0073] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.
Claims
1. A video ringback tone playing method based on off-hook duration prediction algorithm, characterized in that: The steps include: S1, data acquisition step, obtains all necessary data feature information from the user data background and outputs it in CSV data format; S2. Complete quality control on the acquired user feature data, solve the problems of missing values, abnormal values, and no data in the user data information through traditional data cleaning algorithms, and obtain user data information that has passed quality control; S3, screening the key features of the acquired quality-controlled user data, and extracting the corresponding key feature data as regression fitting data; S4, calculation of the called user off-hook duration prediction algorithm; S5, extracting and playing the highlights of the video ringback tone according to the predicted off-hook time of the called user; The steps of the algorithm for predicting the off-hook duration of the called user in S4 are as follows: S41. Through years of experience in communications and observation and analysis of big data, we screened the key feature information of users’ off-hook duration and established a multi-dimensional feature dataset for off-hook duration prediction. At the same time, some semantic features need to be converted into corresponding digital features. S42, label the established user multi-dimensional feature data set respectively, so that each feature vector has a corresponding off-hook duration true value label, and the data set contains features of four dimensions and corresponds to a true value of the call connection duration; S43, the data set is used as the input of the Lasso Regression model. After training and learning the model and updating the parameters, the final model parameters are obtained, and the regression prediction result is obtained. The regression result is the prediction result of the user's call connection time. The regression model is a trained parameter model obtained from the multidimensional feature data. The regression error formula of Lasso Regression is as follows: S44. Send the test set to the trained model for prediction, and evaluate it through the mean square error MSE, root mean square error RMSE, or R-square coefficient to determine the model regression effect. The formula is as follows: The specific steps of S5 are as follows: S51, by splitting the short video into frames, using a dimensionality reduction algorithm to reduce the data dimension of each frame and then putting it into an array, a two-dimensional feature vector representing the entire video data is obtained, and a dimension representing time is added to the feature vector according to the timing information of each frame to obtain a three-dimensional feature vector, and the three-dimensional feature vector represents the feature information of the short video; S52, based on the k-means algorithm, i.e., the K-means algorithm, according to the time feature, the three-dimensional feature vector is clustered unsupervisedly on the basis of retaining the time series information, and the number of clustering categories is also adaptive, so as to extract multiple video clips, which are the highlights of the short video; S53. Based on the predicted duration of the user's call, the video clips are spliced in time sequence, and a wonderful video of no less than the most recently predicted duration is integrated according to the duration of each video clip. By compressing and extracting the video ringback tone content, the user who makes the call can obtain as much video information as possible in a limited time.
2. The video ringback tone playing method based on off-hook duration prediction algorithm according to claim 1, characterized in that: The specific steps of S1 are as follows: S11. First, open the database query window and output the Structured Query Language statement of the data to be exported; S12. Copy all the queried data in the search results and save the results as a Comma-SeparatedValues CSV text file and export it.
3. The video ringback tone playing method based on off-hook duration prediction algorithm according to claim 1, characterized in that: The method for S2 to obtain the quality-controlled user data information is as follows: S21, using neighbor interpolation and average interpolation methods to solve the missing values, neighbor interpolation is to use adjacent data values for interpolation, and average interpolation is to use the average value of the dimension feature for interpolation; S22, using a self-designed threshold to determine abnormal values, the maximum length of a phone call is 54 seconds, if it exceeds the threshold, it is an abnormal value, and negative values and abnormal values are also abnormal values; S23. Filling the call connection duration of the users without data with the average or median of the connection duration of all users.
4. The video ringback tone playing method based on off-hook duration prediction algorithm according to claim 1, characterized in that: The key feature data in S3 include:
1. Off-hook time; 2. Time of day; 3. Gender; 4. Monday to weekend; 5. True value of the user's call connection time.
5. The video ringback tone playing method based on off-hook duration prediction algorithm according to claim 1, characterized in that: The steps of the k-means algorithm are as follows: 1) Evenly select n cluster center points in the time series, where n is one third of the total number of frames; 2) Calculate the average value in the neighborhood of each cluster center point t, where t is half of the distance between the center points; 3) If the absolute value of the difference between the average values of adjacent clusters is less than the set threshold T, the two clusters are merged; 4) Iterate and repeat steps 2) and 3) until no new clusters are generated; 5) According to the number of sample points in each cluster, the top k clusters are taken as the highlights of the short video.
