User VIP prediction method and device in film and television recommendation system and storage medium

Abstract

The invention relates to a movie and television recommendation technology, and discloses a user VIP prediction method and device in a movie and television recommendation system and a storage medium, which can predict a VIP value of a non-VIP user in the recommendation system so as to carry out differentiated recommendation. According to the scheme, the method comprises the following steps: firstly, acquiring original data such as existing VIP data, film watching data, a film and television VIP value and film and television playing amount data of a user; constructing VIP sample data according to the original data; then, correlating the related data, and calculating a feature set of the user; then, according to the feature set of the user, a series of feature values are calculated through feature engineering; thirdly, important feature values are screened out from the series of feature values through a machine learning feature screening method, and a sample data set is constructed; then, the sample data set is divided into a training set and a test set according to a set proportion to complete training of the user VIP prediction model, and finally, the trained user VIP prediction model is utilized to predict the VIP value of the non-VIP user in the recommendation system. The method is suitable for a third-party display platform and can also be migrated to other film and television platforms.

Description

Technical Field

[0001] The present invention relates to film and television recommendation technology, and in particular to a method, device and storage medium for predicting user VIP in a film and television recommendation system. Background Art

[0002] Nowadays, online film and television platforms are one of the important entertainment venues. Users can watch a large number of TV series, movies, variety shows and other programs on the platform. Users on the film and television platform are divided into VIP users and non-VIP users; in comparison, VIP users can watch more and more complete films and televisions, while non-VIP users can only watch free film and television content.

[0003] At present, the film and television recommendation system in traditional technology does not predict the VIP value (the probability of becoming a VIP user) of non-VIP users, and thus cannot make differentiated recommendations for non-VIP users. For example, for potential customers among non-VIP users, they prefer to be recommended film and television resources with content that is more in line with their preferences and has higher popularity; and for non-potential customers among non-VIP users, they prefer to be recommended free film and television content that can be watched in its entirety. Summary of the invention

[0004] The technical problem to be solved by the present invention is to propose a user VIP prediction method, device and storage medium in a film and television recommendation system, which can predict the VIP value of non-VIP users in the recommendation system so as to make differentiated recommendations.

[0005] The technical solution adopted by the present invention to solve the above technical problems is:

[0006] On the one hand, the present invention provides a method for predicting a user VIP in a film and television recommendation system, which is used to predict the VIP value of a non-VIP user in the recommendation system. The method comprises the following steps:

[0007] S1. Obtaining original data: including user's VIP data, movie viewing data, movie VIP value and movie playback data;

[0008] S2. Construct VIP sample data based on the original data;

[0009] S3, associating the VIP sample data with the user's movie viewing data, the movie VIP value and the movie playback volume data, and calculating the user's feature set;

[0010] S4. Based on the user's feature set, a series of feature values ​​are calculated using feature engineering;

[0011] S5. Screening out important feature values ​​from the series of feature values ​​by a machine learning feature screening method to construct a sample data set;

[0012] S6. Divide the sample data set into a training set and a test set according to a set ratio;

[0013] S7, completing the training of the user VIP prediction model according to the training set and the test set;

[0014] S8. Use the trained user VIP prediction model to predict the VIP value of non-VIP users in the recommendation system.

[0015] Furthermore, in step S1, the user's VIP data includes the expiration time of the VIP purchased by the user and whether it has expired; the film and television VIP value indicates whether the corresponding film and television is a paid or free resource.

[0016] Further, in step S2, the VIP sample data includes positive sample VIP data and negative sample VIP data;

[0017] The method for constructing the positive sample data includes: calculating each VIP validity time interval of all users as the positive sample VIP data through the expiration field and VIP expiration time in the user's VIP data;

[0018] The method for constructing negative sample VIP data includes: selecting a time interval of each user that is completely non-overlapping with the positive sample VIP data and has no VIP for at least 30 consecutive days as the negative sample VIP data.

[0019] Furthermore, in step S3, the method of associating the VIP sample data with the user's movie viewing data, the movie VIP value and the movie playback volume data, and calculating the feature set of each user includes:

[0020] First, the VIP sample data is associated with the movie viewing data through the user ID. The movie viewing data is selected in the past 90 days to obtain the movie viewing data set;

[0021] Then, the VIP sample data and the movie viewing data set are intersected, and the intersection time length must be at least 30 days. If the intersection time length is longer than 30 days, only the latest 30 days of movie viewing data are retained to obtain the VIP sample movie viewing set;

[0022] Finally, the VIP sample viewing set is associated with the film and television VIP value and film and television playback data through the film and television ID, and the maximum playback value in the past 30 days is calculated to form the user's feature set.

[0023] Furthermore, in step S4, the series of features include: grouping the user's feature set according to the user and VIP expiration time, calculating the user's viewing features during VIP time and non-VIP time respectively, and viewing features calculated based on the user's viewing of VIP films and TV shows released during the user's VIP sample data time interval.

[0024] Furthermore, in step S7, the user VIP prediction model adopts the LightGBMClassifier model.

[0025] Furthermore, in step S8, when using the trained user VIP prediction model to predict the VIP value of a non-VIP user in the recommendation system, its input is a set of important features calculated and screened based on the movie viewing data of the non-VIP user.

[0026] In a second aspect, the present invention also provides a device for predicting user VIP in a film and television recommendation system, which includes a processor and a memory, wherein the memory stores a computer program, and the processor implements the steps of the method for predicting user VIP in the above-mentioned film and television recommendation system by executing the computer program.

[0027] In a third aspect, the present invention further provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method for predicting user VIP in the above-mentioned film and television recommendation system.

[0028] The beneficial effects of the present invention are:

[0029] The present invention uses machine learning algorithm modeling based on the user's existing VIP data, movie viewing data, film and television VIP value and film and television playback data, so as to predict the VIP value of non-VIP users, so that the film and television recommendation system can make differentiated film and television content recommendations for non-VIP users based on the VIP prediction value, improve user experience, and also promote potential customers to convert into VIP users. This solution is applicable to third-party display platforms and can also be migrated to other film and television platforms. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is a flow chart of a method for predicting user VIP in a film and television recommendation system in an embodiment of the present invention. DETAILED DESCRIPTION

[0031] The present invention aims to propose a user VIP prediction method, device and storage medium in a film and television recommendation system, which can predict the VIP value of non-VIP users in the recommendation system so as to make differentiated recommendations. The present invention obtains the user's real VIP time dimension based on the user VIP data of the existing platform, including the VIP time interval and the non-VIP time interval. Then, by associating the user's movie viewing data (the time interval is set to the past 1 month), the user's movie viewing feature values ​​such as viewing time and type are calculated to form the user's movie viewing feature set. Finally, the model is trained using a machine learning algorithm to complete the user's movie and television VIP prediction.

[0032] The scheme of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0033] Example

[0034] See also Figure 1 The user VIP prediction method in the film and television recommendation system provided in this embodiment includes the following steps:

[0035] (1) Obtaining data:

[0036] In this step, the user's VIP data, movie viewing data, film and television VIP value, and film and television playback data are obtained through the data platform. Among them, the user's VIP data records the validity of the VIP, including the expiration time of the VIP purchased by the user and whether it has expired; the film and television VIP value indicates the relationship between the film and television and the resource party, whether it is paid or free; the movie viewing data can be recorded by the user when watching the movie on the TV terminal using the collection protocol; the film and television playback data can indicate the recent popularity of the film and television, which can be calculated through the viewing data.

[0037] (2) Construct VIP sample data:

[0038] In this step, the VIP sample data includes positive samples and negative samples, where positive samples indicate VIP attributes, while negative samples indicate no VIP attributes. Positive samples can be obtained through VIP expiration marks, and negative samples need to exclude the positive sample time interval of the same user, because users may purchase VIP multiple times, some of which have expired and some have not expired. The time interval of sample data is at least 30 days.

[0039] Specifically, we first calculate the valid time interval of each VIP of all users through the expiration field and the VIP expiration time, and use this as the positive sample VIP data. For the VIP negative sample data, we select a time interval that does not overlap with the VIP positive sample, that is, the same user must completely avoid the positive sample interval in the negative sample interval and must not have VIP for at least 30 consecutive days.

[0040] (3) Correlate user viewing data, film and television VIP value, and film and television playback data:

[0041] In this step, since the movie viewing data is mainly considered when constructing the user VIP prediction model, when constructing the basic features, it is necessary to associate all the related data to form a complete data record. The specific way of association is as follows:

[0042] In step (2), the VIP sample data is associated with the viewing data through the user ID. The viewing data in the past 90 days is selected and set as the viewing set Set(View); then, the intersection of the VIP sample data and Set(View) is calculated. The intersection time length must be at least 30 days. If the intersection time length is longer than 30 days, only the latest 30 days of viewing data are retained to obtain the VIP sample viewing set Set(VIP_V). Next, Set(VIP_V) is associated with the VIP value of the film and television through the film and television id, the film and television playback volume data, and the maximum playback volume value in the past 30 days is calculated, and finally the user's feature set Set(R) is formed.

[0043] (4) Feature Engineering:

[0044] In this step, by calculating the user's movie viewing data, a series of calculated feature items are obtained, including the number of movies watched, the number of episodes, the duration, the number of days, etc. Then the data normalization method is used to normalize the data to the interval [0, 1]. The specific calculation process is as follows:

[0045] For the feature set Set(R) obtained in step (3), group them by user and VIP expiration time, and calculate the following field values ​​of the user during VIP time and non-VIP time respectively:

[0046] How many movies and TV series have you watched in total num, how many episodes of movies and TV series have you watched in total episodes, how many types of movies and TV series are there n_model, how many days of watching movies view_days, how long you have watched movies view_times, the most recent time of watching movies recent_view, the average playback value of movies and TV series play_count, how many VIP movies and TV series have you watched in total num_vip, how many episodes of VIP movies and TV series have you watched episodes_vip, how many types of VIP movies and TV series are there n_model_vip, how many days of watching VIP movies and TV series view_days_vip, how long you have watched VIP movies and TV series view_times_vip, the most recent time of watching VIP movies and TV series recent_view_vip, the average playback value of watching VIP movies and TV series play_count_vip, how many VIP children's movies have you watched in total n um_vip_children, total number of VIP children episodes watched episodes_vip_children, number of days for watching VIP children by type view_days_vip_children, total viewing time of VIP children by type view_times_vip_children, average number of plays of VIP children played by play_count_vip_children, total number of TV series watched num_tv, total number of TV series episodes watched episodes_tv, total number of movies watched num_movie, total number of variety shows watched num_variety, total number of cartoons watched num_cartoon, total number of children's movies watched num_children, total number of children's episodes watched episodes_children.

[0047] The features of VIP movies released during the user's VIP sample data time interval are calculated based on the user's viewing behavior:

[0048] How many VIP TV series i_num_tv have you watched, how many episodes of VIP TV series i_episodes_tv have you watched, the total viewing time of VIP TV series view_times_i_tv, the average playback value of VIP TV series play_count_i_tv, how many VIP movies i_num_movie have you watched, the total viewing time of VIP movies view_times_i_movie, the average playback value of VIP movies play_count_i_movie, how many VIP variety shows i_num_variety have you watched, the total viewing time of VIP variety shows view_times_i_variety, the average playback value of VIP movies play_count_i_movie The average playback value of IP variety shows play_count_i_variety, the number of VIP cartoons watched i_num_cartoon, the total viewing time of VIP cartoons view_times_i_cartoon, the average playback value of watching VIP cartoons play_count_i_cartoon, the total number of VIP children watched i_num_children, the total number of VIP children episodes watched i_episodes_children, the total viewing time of VIP children view_times_i_children, and the average playback value of watching VIP children play_count_i_children.

[0049] In the above features, the most recent viewing time is the most recent viewing time within the time interval corresponding to the VIP sample data. This feature value is a time string, and the number of days between the maximum value of the time interval needs to be calculated.

[0050] Since the distribution of the above eigenvalues ​​varies greatly, for example, the number of movies watched is generally from single digits to dozens, the playback volume of movies is generally from thousands to hundreds of thousands, and the viewing time is generally from single digits to hundreds of hours. Therefore, the normalization method is used to scale the different eigenvalues ​​to the interval [0,1] for subsequent processing. For example, the eigenvalues ​​of the number of movies watched are all divided by 100, the eigenvalues ​​of the playback volume are all divided by 600,000, and the eigenvalues ​​of the total viewing time are all divided by 720 (720 hours in 30 days).

[0051] (5) Screening characteristics:

[0052] Since there are many initial features obtained in step (4), but not every feature has a significant effect on modeling, this step selects features with larger importance coefficients through machine learning feature screening methods, such as the number of movies watched, the number of TV series watched during the VIP time period, etc., which can reduce the amount of model calculations and improve processing efficiency.

[0053] The specific means of screening features are as follows:

[0054] First, the multiple features obtained in step (4) are combined into a vector column through Spark's data converter VectorAssembler. Then, the LightGBMClassifier model of Spark ML (i.e., LightGBM algorithm) is used to set the decision tree depth to 6, the number of iterations to 100, and the learning rate to 0.1. Finally, the fit method is used to complete the feature importance calculation.

[0055] model.getFeatureImportances can output the weight coefficient of each feature. After experiments, the more important features are:

[0056] num, recent_view_diff, play_count, num_movie, num_cartoon, num_children, episodes_children, num_vip, episodes_vip, view_days_vip, play_count_vip, i_num_tv, i_episodes_tv, view_t imes_i_tv, play_count_i_tv, i_num_movie, play_count_i_movie, play_count_i_variety, i_num_cartoon, play_count_i_cartoon, i_num_children, i_episodes_children, view_times_i_chi ldren, num_vip_children, play_count_i_children, play_count_vip_children.

[0057] After the above-mentioned feature screening, the positive and negative sample data of users' movie watching with the above-mentioned important feature values ​​are obtained to construct a sample data set.

[0058] (6) Divide the training set and test set:

[0059] In this step, for the sample data set in step (5), the positive and negative sample data are randomly shuffled, and 50% of each is taken to form the training set and the test set; if the number of positive and negative samples is unbalanced, it can be balanced by downsampling or oversampling. Among them, the training set is used to train the model, and the test set is used to evaluate the performance of the model.

[0060] (7) Training model:

[0061] In this step, the model is trained and its performance is evaluated based on the training set and test set divided in step (6). In an exemplary solution, the machine learning algorithm LightGBM is used to complete the model training. The training method is the Pipeline pipeline, which includes feature processing and model parameter setting. The model effect is evaluated by the AUC value (Area Under Curve).

[0062] (8) Predict user VIP:

[0063] In this step, based on the VIP prediction model trained in step (7), the VIP value of non-VIP users in the recommendation system can be predicted in actual scenarios. In actual scenarios, after obtaining the movie viewing data of non-VIP users, the movie viewing data is processed into the feature set in step (5) as the input of the model. After the calculation of the model, the VIP prediction value of the corresponding user can be obtained, which can reflect the probability of the user becoming a VIP user.

[0064] According to the above process, after obtaining the user's VIP prediction value, the recommendation system can make differentiated recommendations based on the VIP prediction value. For example, for non-VIP users with high VIP prediction values ​​(greater than the threshold), they can be determined as potential customers, and the recommendation system can recommend all VIP movies to them, so that they can understand the movies and TV content that is more in line with their preferences, has a higher current popularity, and has newer content, thereby prompting them to convert into VIP users. For non-VIP users with low VIP prediction values, all free movies and TV can be recommended to them, so that they can watch the movie and TV content in its entirety and improve their viewing experience. A small number of VIP movies and TV and most of the free movies and TV can also be cross-recommended.

[0065] Finally, it should be noted that the above embodiments are only preferred implementations and are not intended to limit the present invention. It should be pointed out that for those skilled in the art, several modifications, equivalent replacements, improvements, etc. can be made without departing from the scope of the present invention and the scope of protection of the claims, and all of these should be included in the protection scope of the present invention.

Claims

1. A user VIP prediction method in a film and television recommendation system, used to predict the VIP value of a non-VIP user in the recommendation system, characterized in that: The method comprises the following steps: S1. Obtaining original data: including user's VIP data, movie viewing data, movie VIP value and movie playback data; S2. Construct VIP sample data based on the original data; S3, associating the VIP sample data with the user's movie viewing data, the movie VIP value and the movie playback volume data, and calculating the user's feature set; S4. Based on the user's feature set, a series of feature values ​​are calculated using feature engineering; S5. Screening out important feature values ​​from the series of feature values ​​by a machine learning feature screening method to construct a sample data set; S6. Divide the sample data set into a training set and a test set according to a set ratio; S7, completing the training of the user VIP prediction model according to the training set and the test set; S8. Use the trained user VIP prediction model to predict the VIP value of non-VIP users in the recommendation system.

2. The method for predicting user VIP in a film and television recommendation system according to claim 1, characterized in that: In step S1, the user's VIP data includes the VIP expiration time and whether it has expired field of a certain VIP purchased by the user; the film and television VIP value indicates whether the corresponding film and television is a paid or free resource.

3. The method for predicting user VIP in a film and television recommendation system according to claim 2, characterized in that: In step S2, the VIP sample data includes positive sample VIP data and negative sample VIP data; The method for constructing the positive sample data includes: calculating each VIP validity time interval of all users as the positive sample VIP data through the expiration field and VIP expiration time in the user's VIP data; The method for constructing negative sample VIP data includes: selecting a time interval of each user that is completely non-overlapping with the positive sample VIP data and has no VIP for at least 30 consecutive days as the negative sample VIP data.

4. The method for predicting user VIP in a film and television recommendation system according to claim 3, characterized in that: In step S3, the method of associating the VIP sample data with the user's movie viewing data, the movie VIP value and the movie playback volume data, and calculating the feature set of each user includes: First, the VIP sample data is associated with the movie viewing data through the user ID. The movie viewing data is selected in the past 90 days to obtain the movie viewing data set; Then, the VIP sample data and the movie viewing data set are intersected, and the intersection time length must be at least 30 days. If the intersection time length is longer than 30 days, only the latest 30 days of movie viewing data are retained to obtain the VIP sample movie viewing set; Finally, the VIP sample viewing set is associated with the film and television VIP value and film and television playback data through the film and television ID, and the maximum playback value in the past 30 days is calculated to form the user's feature set.

5. The method for predicting user VIP in a film and television recommendation system according to claim 1, characterized in that: In step S4, the series of features include: grouping the user's feature set according to the user and VIP expiration time, calculating the user's viewing features during VIP time and non-VIP time respectively, and viewing features calculated based on the user's viewing of VIP films and TV shows released during the user's VIP sample data time interval.

6. The method for predicting user VIP in a film and television recommendation system according to claim 1, characterized in that: In step S7, the user VIP prediction model adopts the LightGBMClassifier model.

7. The method for predicting user VIP in a film and television recommendation system according to claim 1, characterized in that: In step S8, when the trained user VIP prediction model is used to predict the VIP value of a non-VIP user in the recommendation system, its input is a set of important features calculated and screened based on the movie viewing data of the non-VIP user.

8. A device for predicting user VIP in a film and television recommendation system, comprising a processor and a memory, wherein a computer program is stored in the memory, characterized in that: The processor implements the steps of the method for predicting user VIP in a film and television recommendation system according to any one of claims 1 to 7 by executing the computer program.

9. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for predicting user VIP in a film and television recommendation system according to any one of claims 1 to 7 are implemented.

Images