Method and device for recommending paid songs, medium and computing equipment

Abstract

The embodiment of the invention provides a paid song recommendation method and device, a medium and computing equipment, and aims to comprehensively consider information such as song listening preference and payment behavior of historical paid songs under natural distribution and enhanced distribution of a user, more accurately grasp song listening habits and purchase tendency of the user, and improve the user experience. The enhanced distribution result more accurately simulates the distribution state of the paid songs in a real scene, so that the recommendation accuracy of the enhanced distribution mode after the full-song audition link is introduced is improved.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of data analysis, and more specifically, the embodiments of the present invention relate to a method, apparatus, medium and computing device for recommending paid songs. Background Art

[0002] This section is intended to provide a background or context to embodiments of the invention that are recited in the claims. No description herein is admitted to be prior art by inclusion in this section.

[0003] In the current digital music market, the song distribution mechanism affects the operational efficiency and user experience of the music platform. At present, song distribution technologies are mainly divided into natural distribution and enhanced distribution. Among them, natural distribution mainly makes personalized recommendations for non-paid songs based on users' listening preferences, while enhanced distribution recommends paid songs to target user groups according to specific promotion goals, such as increasing purchase rates or increasing exposure, in order to improve economic benefits.

[0004] When the full song trial link is introduced, that is, natural distribution can recommend paid songs that support full song free trial, for paid songs that support full song free trial, their real data distribution includes two parts: natural distribution and enhanced distribution. The current enhanced distribution continues to generate enhanced distribution results according to the logic that natural distribution does not support the distribution of paid songs that are free to listen to the whole song, ignoring the impact of the distribution of paid songs under natural distribution on the enhanced distribution effect, thereby affecting the accuracy of the recommendation results of enhanced distribution. Summary of the invention

[0005] In view of this, in order to solve the above problems, the embodiments of the present invention are intended to provide a method, apparatus, medium and computing device for recommending paid songs.

[0006] In a first aspect of the embodiments of the present invention, a method for recommending paid songs is provided, the method comprising:

[0007] Acquire music-related features of the user to be recommended under natural distribution and enhanced distribution; the music-related features at least include song preference features and payment behavior features of the user to be recommended for paid songs; the natural distribution is used to recommend songs to the user according to the user's song preference information; the enhanced distribution is used to recommend paid songs to the user according to the set business goals;

[0008] According to the music-related features and the song features of multiple candidate paid songs in the resource library, the distribution probability of the candidate paid songs is predicted; the candidate paid songs support free trial listening of the entire song;

[0009] According to the distribution probability of the candidate paid songs, a set number of candidate paid songs are selected from the plurality of candidate paid songs as song recommendation results for enhanced distribution.

[0010] Optionally, the method further comprises:

[0011] Obtain song recommendation results for the user to be recommended under natural distribution;

[0012] The song recommendation results of the enhanced distribution are incorporated into the song recommendation results of the natural distribution to obtain the target recommendation results for the user to be recommended.

[0013] Optionally, predicting the distribution probability of the candidate paid songs according to the music-related features and song features of a plurality of candidate paid songs in a resource library includes:

[0014] Inputting the music-related features and the song features of the candidate paid songs into a song distribution network to obtain the distribution probability of the candidate paid songs;

[0015] The pre-trained samples of the song distribution network include: music-related features of sample users; song features of historical paid songs of sample users under natural distribution and enhanced distribution;

[0016] The true label of the pre-trained sample is used to indicate whether the user purchased the historical paid song.

[0017] Optionally, the step of inputting the music-related features and the song features of the candidate paid songs into a song distribution network to obtain the distribution probability of the candidate paid songs includes:

[0018] Input the music-related features and the song features of the candidate paid songs into the natural distribution expert network, the enhanced distribution expert network and the scene sharing network respectively to obtain the corresponding first feature vector, the second feature vector and the third feature vector;

[0019] Obtaining a contribution weight according to the first eigenvector and the second eigenvector;

[0020] Multiplying the contribution weight by the first eigenvector to obtain a weighted first eigenvector;

[0021] Obtaining the distribution probability of the candidate paid song according to the second feature vector, the third feature vector and the weighted first feature vector;

[0022] Among them, the natural distribution expert network is used to extract features related to natural distribution; the enhanced distribution expert network is used to extract features related to enhanced distribution; and the scene sharing network is used to extract features common to both natural distribution and enhanced distribution.

[0023] Optionally, the method further comprises:

[0024] Obtaining the domain statistical features of the user to be recommended under natural distribution and enhanced distribution; the domain statistical features are used to represent the statistical description of the interactive behavior, and the interactive behavior refers to the interactive behavior between the user to be recommended and the paid songs recommended in the past;

[0025] The domain statistical features, the music-related features and the song features of the candidate paid songs are simultaneously input into the natural distribution enhancement network to obtain the corresponding first feature vector.

[0026] Optionally, obtaining a contribution weight according to the first eigenvector and the second eigenvector includes:

[0027] Obtaining an inner product operation result of the first eigenvector and the second eigenvector; determining the contribution weight according to the inner product operation result;

[0028] Alternatively, similarity calculation is performed through a network layer based on the first feature vector and the second feature vector, and the contribution weight is determined according to the similarity calculation result.

[0029] Optionally, the method further comprises a pre-training step of the song distribution network:

[0030] Input the song features of the historical paid songs and the corresponding music-related features into the natural distribution expert network, the enhanced distribution expert network and the scene sharing network respectively to obtain the corresponding fourth feature vector, fifth feature vector and sixth feature vector;

[0031] Obtaining a sample weight according to the fourth eigenvector and the fifth eigenvector;

[0032] Obtaining the predicted distribution probability of the historical paid songs according to the sample weight, and the obtained fourth eigenvector, fifth eigenvector, and sixth eigenvector;

[0033] A loss function is calculated based on the predicted distribution probability of the historical paid songs and the true labels of the historical paid songs, and the parameters of the song distribution network are adjusted inversely to minimize the loss function.

[0034] Optionally, obtaining the predicted distribution probability of the historical paid songs according to the sample weight, and the obtained fourth eigenvector, fifth eigenvector, and sixth eigenvector includes:

[0035] In the case where the historical paid songs belong to samples under natural distribution, the sample weight is multiplied by the fifth eigenvector to obtain a seventh eigenvector;

[0036] The predicted distribution probability of the historical paid songs is obtained according to the fourth eigenvector, the sixth eigenvector and the seventh eigenvector.

[0037] Optionally, obtaining the predicted distribution probability of the historical paid songs according to the sample weight, and the obtained fourth eigenvector, fifth eigenvector, and sixth eigenvector includes:

[0038] In the case where the historical paid song belongs to a sample under enhanced distribution, multiplying the sample weight by the fourth feature vector to obtain an eighth feature vector;

[0039] The predicted distribution probability of the historical paid songs is obtained according to the fifth eigenvector, the sixth eigenvector and the eighth eigenvector.

[0040] Optionally, the pre-training step further includes:

[0041] For each historical paid song in the pre-trained sample, obtain the song status of the historical paid song; the song status includes at least one of free trial listening of the whole song, free trial listening of a fragment, and trial listening not supported;

[0042] Inputting the payment status of the historical paid songs into the correction network at the same time, and obtaining the correction score output by the correction network;

[0043] The predicted distribution probability of the historical paid songs is fused with the correction score to obtain an updated distribution probability.

[0044] In a second aspect of the embodiments of the present invention, a device for recommending paid songs is provided, the device comprising:

[0045] A feature acquisition module is used to acquire music-related features of the user to be recommended under natural distribution and enhanced distribution; the music-related features at least include song preference features and payment behavior features of the user to be recommended for paid songs; the natural distribution is used to recommend songs to the user according to the user's song preference information; the enhanced distribution is used to recommend paid songs to the user according to the set business goals;

[0046] A prediction module, used to predict the distribution probability of the candidate paid songs according to the music-related features and the song features of multiple candidate paid songs in the resource library; the candidate paid songs support free trial listening of the entire song;

[0047] The enhanced distribution result acquisition module is used to select a set number of candidate paid songs from the multiple candidate paid songs as song recommendation results for enhanced distribution according to the distribution probability of the candidate paid songs.

[0048] Optionally, the device further comprises:

[0049] Obtain song recommendation results for the user to be recommended under natural distribution;

[0050] The song recommendation results of the enhanced distribution are incorporated into the song recommendation results of the natural distribution to obtain the target recommendation results for the user to be recommended.

[0051] Optionally, the prediction module is specifically used for:

[0052] Inputting the music-related features and the song features of the candidate paid songs into a song distribution network to obtain the distribution probability of the candidate paid songs;

[0053] Among them, the pre-trained samples of the song distribution network include: music-related features of sample users; song features of historical paid songs of sample users under natural distribution and enhanced distribution; and the true labels of the pre-trained samples are used to indicate whether the user purchased the historical paid songs.

[0054] Optionally, the prediction module is specifically used for:

[0055] The music-related features and the song features of the candidate paid songs are respectively input into the natural distribution expert network, the enhanced distribution expert network and the scene sharing network to obtain the corresponding first feature vector, the second feature vector and the third feature vector; according to the first feature vector and the second feature vector, a contribution weight is obtained; the contribution weight is multiplied by the first feature vector to obtain a weighted first feature vector; according to the second feature vector, the third feature vector and the weighted first feature vector, a distribution probability of the candidate paid song is obtained;

[0056] Among them, the natural distribution expert network is used to extract features related to natural distribution; the enhanced distribution expert network is used to extract features related to enhanced distribution; and the scene sharing network is used to extract features common to both natural distribution and enhanced distribution.

[0057] Optionally, the device further comprises:

[0058] Obtaining the domain statistical features of the user to be recommended under natural distribution and enhanced distribution; the domain statistical features are used to represent the statistical description of the interactive behavior, and the interactive behavior refers to the interactive behavior between the user to be recommended and the paid songs recommended in the past;

[0059] The domain statistical features, the music-related features and the song features of the candidate paid songs are simultaneously input into the natural distribution enhancement network to obtain the corresponding first feature vector.

[0060] Optionally, when the prediction module is used to obtain a contribution weight according to the first eigenvector and the second eigenvector, the prediction module includes:

[0061] Obtaining an inner product operation result of the first eigenvector and the second eigenvector; determining the contribution weight according to the inner product operation result;

[0062] Alternatively, similarity calculation is performed through a network layer based on the first feature vector and the second feature vector, and the contribution weight is determined according to the similarity calculation result.

[0063] Optionally, the device further comprises a pre-training step of the song distribution network:

[0064] A sample input module, used to input the song features and corresponding music-related features of the historical paid songs into the natural distribution expert network, the enhanced distribution expert network and the scene sharing network, respectively, to obtain the corresponding fourth feature vector, fifth feature vector and sixth feature vector;

[0065] A weight determination module, used for obtaining a sample weight according to the fourth eigenvector and the fifth eigenvector;

[0066] A sample prediction module, used to obtain the predicted distribution probability of the historical paid songs according to the sample weight, and the obtained fourth eigenvector, fifth eigenvector and sixth eigenvector;

[0067] The loss function calculation module is used to calculate the loss function according to the predicted distribution probability of the historical paid songs and the real labels of the historical paid songs, and reversely adjust the parameters of the song distribution network to minimize the loss function.

[0068] Optionally, the sample prediction module is specifically used for:

[0069] In the case that the historical paid song is a sample under natural distribution, the sample weight is multiplied by the fifth eigenvector to obtain the seventh eigenvector; based on the fourth eigenvector, the sixth eigenvector and the seventh eigenvector, the predicted distribution probability of the historical paid song is obtained.

[0070] Optionally, the sample prediction module is specifically used for:

[0071] In the case where the historical paid song belongs to the sample under enhanced distribution, the sample weight is multiplied by the fourth eigenvector to obtain an eighth eigenvector; based on the fifth eigenvector, the sixth eigenvector and the eighth eigenvector, the predicted distribution probability of the historical paid song is obtained.

[0072] Optionally, the pre-training step further includes:

[0073] For each historical paid song in the pre-trained sample, obtain the song status of the historical paid song; the song status includes at least one of free trial listening of the whole song, free trial listening of a fragment, and trial listening not supported;

[0074] Inputting the payment status of the historical paid songs into the correction network at the same time, and obtaining the correction score output by the correction network;

[0075] The predicted distribution probability of the historical paid songs is fused with the correction score to obtain an updated distribution probability.

[0076] In a third aspect of the embodiments of the present invention, a readable storage medium is provided, on which a computer program is stored, and when the program is executed by a processor, the above-mentioned method for recommending paid songs is implemented.

[0077] In a fourth aspect of the embodiment of the present invention, a computing device is provided, comprising: a processor and a memory; the memory is used to store a computer program; the processor is used to execute the above-mentioned paid song recommendation method by calling the computer program.

[0078] The embodiments of the present invention more accurately grasp the user's listening habits and purchasing tendencies by comprehensively considering the user's listening preferences, payment behaviors, etc. for historical paid songs distributed naturally and historical paid songs distributed in an enhanced manner, and achieve accurate prediction of the user's potential purchasing intention for candidate paid songs. This method of comprehensively considering multiple information sources makes the enhanced distribution results more accurately simulate the distribution status of paid songs in real scenarios. In the process of enhanced distribution with the introduction of the full song audition link, the recommendation accuracy of the enhanced distribution method after the introduction of the full song audition link is improved. While considering the user's personalized needs, the goal of business promotion is taken into account, so that the paid songs finally recommended can not only meet the user's preferences, but also meet the business needs of increasing the song purchase rate, achieving a win-win situation for users and platforms. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] The above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily understood by reading the detailed description below with reference to the accompanying drawings. In the accompanying drawings, several embodiments of the present invention are shown in an exemplary and non-limiting manner, in which:

[0080] Figure 1 A flowchart of a method for recommending paid songs according to an embodiment of the present invention is schematically shown;

[0081] Figure 2 An example diagram schematically shows a song recommendation result integrating natural distribution and enhanced distribution according to an embodiment of the present invention;

[0082] Figure 3 A schematic diagram of a training and reasoning process of a song distribution network according to an embodiment of the present invention is schematically shown;

[0083] Figure 4A A schematic diagram of a network structure of a song distribution network according to an embodiment of the present invention is schematically shown;

[0084] Figure 4B A schematic diagram of an input part of a song distribution network according to an embodiment of the present invention is schematically shown;

[0085] Figure 5A A schematic diagram of a process of training a song distribution network using training samples belonging to a natural distribution method according to an embodiment of the present invention is schematically shown;

[0086] Figure 5B A schematic diagram of a process of training a song distribution network using training samples belonging to an enhanced distribution method according to an embodiment of the present invention is schematically shown;

[0087] Figure 6 A diagram schematically shows an example of a song distribution network including song state correction in a training phase according to an embodiment of the present invention;

[0088] Figure 7 A block diagram schematically shows a device for recommending paid songs according to an embodiment of the present invention;

[0089] Figure 8 A schematic diagram of a medium according to an embodiment of the present invention is schematically shown;

[0090] Fig. 9 A schematic diagram of a computing device according to an embodiment of the present invention is schematically shown.

[0091] In the drawings, the same or corresponding reference numerals represent the same or corresponding parts. DETAILED DESCRIPTION

[0092] The principles and spirit of the present invention will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided only to enable those skilled in the art to better understand and implement the present invention, and are not intended to limit the scope of the present invention in any way. On the contrary, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art.

[0093] Those skilled in the art know that the embodiments of the present invention can be implemented as a system, device, apparatus, method or computer program product. Therefore, the present invention can be specifically implemented in the following forms, namely: complete hardware, complete software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

[0094] According to an embodiment of the present invention, a method, apparatus, medium and computing device for recommending paid songs are proposed.

[0095] It should be understood herein that any number of elements in the drawings is for illustration rather than limitation, and any naming is only for distinction rather than having any limiting meaning.

[0096] The principle and spirit of the present invention are explained in detail below with reference to several representative embodiments of the present invention.

[0097] In the current digital music market, the song distribution mechanism affects the operational efficiency and user experience of the music platform. Currently, song distribution technologies are mainly divided into natural distribution and enhanced distribution.

[0098] Among them, natural distribution mainly recommends non-paid songs in a personalized way based on the user's listening preferences, that is, natural distribution is used to recommend free songs. It recommends free songs to users based on natural factors such as user historical behavior and interest preferences, and does not involve the promotion of paid content. Enhanced distribution recommends paid songs to target user groups according to specific promotion goals, such as increasing purchase rates or exposure, to improve economic benefits. Enhanced distribution is used to promote paid content. It recommends paid songs to potential users through specific recommendation strategies, algorithms or adjusted recommendation results to achieve user payment conversion.

[0099] Because natural distribution does not involve paid content, the distribution logic of enhanced distribution in related technologies is trained based on collected samples of paid songs that have been historically distributed in enhanced distribution scenarios. Both the distribution logic training and services of enhanced distribution focus on the prediction of a single paid task.

[0100] When promoting the full song trial link, that is, when paid content starts to support free full song trials, paid songs that support free full song trials can be recommended based on natural distribution, and enhanced distribution itself can perform extraordinary on paid songs. Therefore, for paid songs that support free full song trials, their real data distribution includes two parts: natural distribution and enhanced distribution. The current enhanced distribution continues to generate enhanced distribution results according to the logic that natural distribution does not support the distribution of paid songs that do not support free full song trials, ignoring the impact of the distribution of paid songs under natural distribution on the enhanced distribution effect, thereby affecting the accuracy of the recommendation results of enhanced distribution.

[0101] Based on this, the present invention proposes a method for recommending paid songs, which introduces information such as users' listening preferences, payment behaviors, etc. for historical paid songs under natural distribution and enhanced distribution methods, so as to more comprehensively and accurately grasp users' listening habits and purchasing tendencies. The method of comprehensively considering multiple information sources makes the enhanced distribution results more accurately simulate the distribution status of paid songs when promoting the full song trial link. In the process of introducing the enhanced distribution of the full song trial link, the personalized needs of users are considered while taking into account the goal of business promotion, so that the paid songs finally recommended can not only meet the preferences of users, but also meet the business needs of increasing the song purchase rate.

[0102] After introducing the basic principles of the present invention, various non-limiting embodiments of the present invention are described in detail below.

[0103] Exemplary Methods

[0104] Reference below Figure 1-9 The method for recommending paid songs according to an exemplary embodiment of the present invention is described below. It should be noted that the above application scenarios are only shown to facilitate understanding of the spirit and principle of the present invention, and the embodiments of the present invention are not limited in this regard. On the contrary, the embodiments of the present invention can be applied to any applicable scenario.

[0105] See also Figure 1 The paid song recommendation method provided in this embodiment may include at least the following steps:

[0106] S101, obtaining music-related features of the user to be recommended under natural distribution and enhanced distribution; the music-related features at least include song preference features and payment behavior features of the user to be recommended for paid songs; the natural distribution is used to recommend songs to the user according to the user's song preference information; the enhanced distribution is used to recommend paid songs to the user according to the set business goals;

[0107] The user to be recommended is a user for whom a song recommendation result including paid songs needs to be generated in the system or platform.

[0108] The basis for the song recommendation of natural distribution is the user's song preference, that is, the user's listening preference is analyzed through the data of the type of songs listened to by the user in history, the singer preference, the favorite songs, the songs liked, etc., so as to recommend songs that meet the user's listening preference. The song distribution objects of the natural distribution described in this embodiment include free songs and paid songs that support full free trial listening. That is, in the process of executing natural distribution to recommend songs for users, multiple songs that meet the user's listening preference are selected from free songs and paid songs that support full free trial listening to generate the song recommendation results of natural distribution, thereby realizing personalized song recommendations based on the user's listening preference.

[0109] Enhanced distribution adjusts song recommendation results through human intervention or specific strategies based on natural distribution to achieve specific business goals, such as increasing the exposure of specific paid songs. Unlike natural distribution, which uses recommendation algorithms to provide users with personalized song recommendations that meet song preferences, enhanced distribution aims to achieve set business goals through strategy optimization, and its song recommendations are guided by business goals. For example, enhanced distribution can forcibly incorporate paid songs into a fixed position in the song recommendation list of natural distribution through specific strategies or human intervention to increase the penetration rate of the incorporated paid songs.

[0110] Based on the method provided in the present application applied to the song distribution scenario under the full song free trial link, before the introduction of the full song free trial link, enhanced distribution is used to distribute all paid songs, including three types of paid songs that support full song free trial, fragment trial and do not support trial. After the introduction of the full song free trial link, enhanced distribution under the full song free trial link is specifically used to distribute paid songs that support full song free trial to generate corresponding recommendation results.

[0111] The music-related features include various features used to construct user music portraits, including music-related features under natural distribution methods and music-related features under enhanced distribution methods. The music-related features may at least include song preference features and payment behavior features of the recommended user for paid songs. Among them, the song preference features are used to indicate the preference tendency and degree of the recommended user for different types of paid songs, reflecting the user's musical taste and aesthetic tendency for paid songs, and can be constructed by analyzing the historical interaction data such as likes, collections, plays, comments, and shares of the recommended user for historical paid songs recommended by natural distribution / enhanced distribution methods; the payment behavior features are used to indicate the user's willingness to purchase different types of paid songs, consumption habits, and payment ability, and can reflect the distribution of paid songs that the user is willing to pay for, that is, which types of paid songs the user tends to pay for, and can be constructed by analyzing the purchase behavior data of the recommended user for historical paid songs recommended by natural distribution / enhanced distribution methods, such as paying for singles and recharging for paid songs.

[0112] The music-related features may also include, but are not limited to, user personal information (such as user name, gender, age, region, etc.), song search history, social information (such as the user's friend list on the song platform, community participation) and device information (such as the type of device used by the user, operating system, screen resolution), album / single purchase information, and user historical sequence data (such as song playback history, collection history, search history, and sharing history).

[0113] S102, predicting the distribution probability of the candidate paid songs according to the music-related features and the song features of multiple candidate paid songs in the resource library; the candidate paid songs support free trial listening of the entire song;

[0114] Among them, the predicted distribution probability is used to represent the willingness of the recommended user to pay for the candidate paid song, which is a value between 0 and 1 and can be calculated based on the degree of match between the user's music-related features and the song features of the candidate song.

[0115] The resource library represents a data set containing a large number of paid songs that support full free trial listening. In order to improve the efficiency and accuracy of the recommendation system, a preset number of paid songs can be selected from the resource library as candidate paid songs through rapid recall.

[0116] The song features of the candidate paid songs are used to comprehensively describe the various attributes of the songs, which may include but are not limited to audio features, lyrics features, metadata features, user interaction features, and audition feedback features. Among them, audio features may include rhythm, pitch, loudness and other relevant dimensional features that describe audio; lyrics features may be constructed from dimensions such as lyrics theme, emotional tendency, and vocabulary richness; metadata features may include artist information, album information, genre, release time, etc.; user interaction features may be constructed from interactive dimensions such as the number of song plays, the number of favorites, the number of comments, and the number of shares; audition feedback features may be constructed from dimensions such as audition completion rate, skip rate, and repeated play rate.

[0117] Regarding predicting the distribution probability of candidate paid songs in the resource library, it can be achieved based on the degree of matching between the music-related features and the song features of the candidate paid songs (such as based on feature similarity or distance measurement), or it can use neural networks or machine learning to use the music-related features and the song features of the candidate paid songs as input to predict the distribution probability.

[0118] S103, selecting a set number of candidate paid songs from the plurality of candidate paid songs as song recommendation results for enhanced distribution according to the distribution probabilities of the candidate paid songs.

[0119] The set number indicates the number of songs to be selected as recommendation results from a plurality of candidate paid songs, and the data may be determined based on factors such as business requirements, the display capability of the user interface and / or the performance of the recommendation system.

[0120] The enhanced distribution song recommendation result refers to a song list screened from candidate paid songs based on distribution probability and displayed to users. The song list can be displayed in order of distribution probability from high to low, or the screened songs can be randomly sorted.

[0121] For multiple candidate paid songs in the resource library, after obtaining the distribution probability corresponding to each candidate paid song, the multiple candidate paid songs can be sorted in descending order according to the distribution probability, and then the first N songs are selected from the sorted candidate paid songs according to the set number N, and the song recommendation results of enhanced distribution are generated based on the first N songs. In the case that the total number of the multiple candidate paid songs is less than the set number, all the multiple candidate paid songs can be selected and the corresponding song recommendation results can be generated.

[0122] In the disclosed embodiment, by introducing a full song trial link, enhanced distribution focuses on distributing paid songs that support full song free trial, and in the process of enhanced distribution, the music-related characteristics of the recommended users under natural distribution and enhanced distribution, including song preference characteristics and payment behavior characteristics, are comprehensively considered, involving multiple information sources to accurately grasp the user's listening habits and purchasing tendencies, so that in the process of predicting the distribution probability of candidate paid songs, the business promotion goals are considered while taking into account the user's personalized song preference needs, so that the paid songs recommended by the enhanced distribution can finally meet the user's preferences, and meet the business needs of increasing the song purchase rate, and meet the needs of improving user satisfaction and loyalty, and promoting the platform's business growth and revenue increase.

[0123] In addition, the song recommendation results of enhanced distribution are all paid songs that support free full-song trials, which meets the user's demand for a complete song experience and improves user satisfaction and engagement through accurate recommendations. Users can listen and decide whether to purchase without any burden, thereby optimizing the user experience.

[0124] In some optional implementations, after the full song trial link is introduced based on the present invention, natural distribution can recommend free songs, and natural distribution and enhanced distribution can also simultaneously recommend paid songs that support full song free trial. Therefore, after obtaining the song recommendation results of the enhanced distribution, the song recommendation results of the user to be recommended under natural distribution can also be obtained; next, the song recommendation results of the enhanced distribution are incorporated into the song recommendation results of the natural distribution to obtain the target recommendation results for the user to be recommended.

[0125] That is, by analyzing the user's historical song playing, liking, collection, sharing and other interactive behavior data or historical sequence data, user personal information, song search history, social information, device information, album / single purchase information and other information, the personalized listening preferences of the user to be recommended can be determined without considering whether the song is paid, and then according to the personalized listening preferences, a preset number of songs are matched from the naturally distributed song resource pool through an algorithm as the naturally distributed song recommendation results. Among them, the songs in the song resource pool include free songs and paid songs that support full free trial listening. Therefore, the naturally distributed song recommendation results include at least one of free songs and paid songs that support full free trial listening.

[0126] After obtaining the song recommendation results of natural distribution and the song recommendation results of enhanced distribution, the song recommendation results of enhanced distribution can be merged into the song recommendation results of natural distribution according to a pre-set result merging method, and the song recommendation results of natural distribution that have been processed by song merging can be used as the target recommendation results for the user to be recommended. For example, the song insertion positions in the song recommendation results of natural distribution can be pre-set, and then the paid songs included in the song recommendation results of enhanced distribution can be sequentially merged into the song insertion positions, so as to achieve the merging of the results of natural distribution and enhanced distribution.

[0127] like Figure 2 An example diagram schematically shows a song recommendation result integrating natural distribution and enhanced distribution. Assume that the target song recommendation result for the user to be recommended sets 30 song display positions, where the songs at the 2nd, 5th, 10th, 15th, and 20th sequence positions are specifically set to be recommended songs obtained through enhanced distribution. Then:

[0128] After obtaining 25 song recommendation results through natural distribution, the recommended songs A, B, C, D, and E obtained through enhanced distribution can be inserted into the 25 song recommendation results in sequence, so that the recommended songs A, B, C, D, and E meet the 2nd, 5th, 10th, 15th, and 20th sequence positions in the target song recommendation results.

[0129] In the process of merging the song recommendation results of natural distribution and enhanced distribution, the song recommendation results of natural distribution and enhanced distribution can be deduplicated first, such as deleting duplicate songs that appear in the song recommendation results of enhanced distribution from the song recommendation results of natural distribution; after deduplication, the song recommendation results in enhanced distribution are inserted into the song recommendation results of natural distribution in sequence. If the number of songs in the song recommendation results of natural distribution is less than the number of songs required by the target song recommendation results after the song insertion process, new songs can be selected as enhanced distribution results according to the distribution probability of candidate paid songs, and inserted into the current song recommendation results of natural distribution after the existing songs are sorted, to make up for the missing number of songs.

[0130] In the disclosed embodiment, after the introduction of the full song trial link, paid songs that support full song free trial can be naturally recommended to users. By incorporating the song recommendation results of enhanced distribution into the song recommendation results of natural distribution, the two distribution methods complement each other. The combination of the two makes the recommendation results more comprehensive and diverse, meets the different listening needs of users, and provides users with more opportunities to try and discover high-quality paid songs. At the same time, paid songs are naturally integrated into the recommendation list, and users can listen to them without additional search or payment, thereby lowering the threshold for obtaining paid songs and improving users' willingness and satisfaction.

[0131] In some optional implementations, in order to more accurately grasp the user's music preferences and accurately analyze the user's willingness to pay for paid songs, for the aforementioned step S102, based on the music-related features and the song features of multiple candidate paid songs in the resource library, predicting the distribution probability of the candidate paid songs, this embodiment provides a song distribution probability prediction method based on a neural network, which can efficiently process and analyze complex music feature data, thereby providing a more accurate distribution probability prediction. The song distribution probability prediction method based on a neural network includes the following implementation process:

[0132] The music-related features and the song features of the candidate paid songs are input into a song distribution network to obtain the distribution probability of the candidate paid songs; wherein the pre-trained samples of the song distribution network include: the music-related features of the sample users, and the song features of the sample users' respective historical paid songs under natural distribution and enhanced distribution, and these features together constitute the input part of the training samples; the true label of the pre-trained sample is used to indicate whether the user has purchased the historical paid song, and can be set to a binary classification label, such as 1 for purchase and 0 for not purchase.

[0133] That is, the song distribution network described in this embodiment is dedicated to predicting distribution probability. In the pre-training stage of the song distribution network, each training sample includes the music-related features of the sample user and the song features of the sample user's historical paid songs, and a true label is set for the training sample to indicate whether the user has purchased the historical paid song. Among them, the music-related features include music-related features under natural distribution and enhanced distribution; and the sample user's historical paid songs include historical paid songs under any distribution method of natural distribution and enhanced distribution. For example, for sample user A, 100 paid songs are obtained through natural distribution and enhanced distribution, respectively, then 200 training samples are formed for sample user A, and each training sample of the example may include the following dimensional representation information [user A's relevant music features, distribution method, song features of historical paid songs, whether to support free trial of the entire song, whether to purchase].

[0134] like Figure 3The schematic diagram of the training and inference process of the song distribution network is shown as an example. During the training phase, the training samples are input into the song distribution network, so that the song distribution network predicts the probability of the user purchasing the input historical paid songs based on the information input into the network, and compares the predicted probability with the real label of the historical paid song, that is, the corresponding true value probability (0 or 1) indicating whether the song is purchased. The difference between the predicted value and the real value is measured by calculating the loss function, and the back propagation algorithm is used to adjust the network parameters in the song distribution network according to the gradient information of the loss function to minimize the loss function. The above process is repeated until the training convergence condition is reached, that is, the loss function value no longer decreases significantly or reaches the preset number of iterations, and the training of the song distribution network is completed.

[0135] The song distribution network learns how to extract effective information from the input music-related features and song features and predict the probability of users purchasing songs during the training phase. Therefore, during the inference phase, the music-related features of the user to be recommended and the song features of the candidate paid songs are input into the trained song distribution network. The song distribution network can use the knowledge learned during the training phase to efficiently process the input features and output the distribution probability of each candidate paid song, reflecting the possibility that the user to be recommended will purchase or be interested in the candidate paid song.

[0136] In the disclosed embodiment, based on the ability of neural networks to learn complex nonlinear relationships, the embodiment utilizes a song distribution network based on a neural network and introduces two types of samples, natural distribution and enhanced distribution, in the training phase to accurately simulate the song distribution distribution of paid songs in the full song audition link, so that the trained song distribution network can efficiently process and analyze input music-related features and song features of candidate paid songs, comprehensively consider multiple information sources to accurately predict the distribution probability of candidate paid songs, and optimize the enhanced distribution of paid song recommendations, thereby improving the accuracy of distribution probability prediction and paid song recommendations under the enhanced distribution method.

[0137] In some alternative embodiments, see Figure 4AA schematic diagram of the network structure of a song distribution network is shown as an example. The feature extraction layer of the song distribution network described in this embodiment may include at least a natural distribution expert network, an enhanced distribution expert network, and a scene sharing network, wherein the natural distribution expert network is used to extract features related to natural distribution; the enhanced distribution expert network is used to extract features related to enhanced distribution; and the scene sharing network is used to extract features common to both natural distribution and enhanced distribution. That is, the natural distribution expert network is used to capture the inherent laws and patterns that conform to the natural distribution method from the input features, while the enhanced distribution expert network is used to capture the inherent laws and patterns under the enhanced distribution method from the input features. The scene sharing network is used to extract shared information between enhanced distribution and natural distribution, representing common features or laws across scenarios, and providing the song distribution network with an in-depth understanding of the similarities and differences between the two distribution methods.

[0138] Based on the structure of the song distribution network, the process of inputting the music-related features and the song features of the candidate paid songs into the song distribution network to obtain the distribution probability of the candidate paid songs as described in the above embodiment can be implemented through the following steps a1-a4:

[0139] a1, inputting the music-related features and the song features of the candidate paid songs into the natural distribution expert network, the enhanced distribution expert network and the scene sharing network respectively, to obtain the corresponding first feature vector, second feature vector and third feature vector;

[0140] The music-related features and the song features of the candidate paid songs are used as the input of the song distribution network, and are simultaneously input into the natural distribution expert network, the enhanced distribution expert network, and the scene sharing network for processing, see Figure 4A As shown, the natural distribution expert network performs feature extraction on the network input and outputs a first feature vector, the enhanced distribution expert network performs feature extraction on the network input and outputs a second feature vector, and the scene sharing network performs feature extraction on the network input and outputs a third feature vector.

[0141] a2, obtaining a contribution weight according to the first eigenvector and the second eigenvector;

[0142] Based on the distribution probability predicted for the candidate paid songs in this embodiment, it is used to enhance the distribution scenario. Therefore, the contribution weight is used to represent the degree of information contribution of the natural distribution related features extracted by the natural distribution expert network to the enhanced distribution scenario, and determines the degree of retention of the natural distribution related features in the current enhanced distribution scenario, so as to better integrate the information from different expert networks and enhance the feature expression.

[0143] In this embodiment, the contribution weight represents the degree of retention of the first feature vector output by the natural distribution expert network in the current enhanced distribution scenario. The correlation or similarity between the features output by the natural distribution expert network and the enhanced distribution expert network can be calculated to identify the importance of the first feature vector output by the natural distribution expert network to the enhanced distribution scenario, and then determine the contribution weight.

[0144] a3, multiplying the contribution weight by the first eigenvector to obtain a weighted first eigenvector;

[0145] Based on the contribution weight used to represent the retention degree of the first feature vector in the current enhanced distribution scenario, therefore, Figure 4A As shown, by multiplying the contribution weight with the first eigenvector output by the natural distribution expert network, a weighted first eigenvector is obtained.

[0146] a4, obtaining the distribution probability of the candidate paid song according to the second feature vector, the third feature vector and the weighted first feature vector;

[0147] like Figure 4A As shown, the second feature vector, the third feature vector and the weighted first feature vector obtained by feature extraction can be fused to obtain a fused feature, and then the distribution probability of the paid song can be predicted based on the fused feature.

[0148] Regarding the acquisition of fusion features, the second eigenvector, the third eigenvector and the weighted first eigenvector can be spliced; or, a learnable weight can be set for each of the three eigenvectors during the training phase, and the learnable weight can be used as a weighting parameter after pre-training. After the second eigenvector, the third eigenvector and the weighted first eigenvector, each eigenvector is multiplied by its corresponding learnable weight to obtain the weighted result of the eigenvector, and then the weighted results of the second eigenvector, the third eigenvector and the weighted first eigenvector are spliced ​​or added to obtain the fusion feature; or, the second eigenvector, the third eigenvector and the weighted first eigenvector can be input into the trained fusion layer to obtain the fusion feature outputted therefrom. In addition, other technical means for realizing the fusion representation of multiple features in the related art can also be used for implementation, and the present invention does not limit this.

[0149] By multiplying the contribution weight with the first eigenvector to obtain a weighted first eigenvector, and fusing it with the second and third eigenvectors, the song distribution network can construct a more refined and accurate feature representation. This fused feature not only includes the natural distribution and enhanced distribution characteristics of the song, but also integrates the shared information across scenarios, thereby improving the prediction accuracy of the distribution probability.

[0150] In the disclosed embodiment, music-related features and song features of candidate paid songs are respectively input into the natural distribution expert network, the enhanced distribution expert network and the scene sharing network, so that the song distribution network can capture feature information of multiple dimensions. The multi-network parallel processing method ensures the diversity and comprehensiveness of feature representation, which facilitates a more accurate understanding of user preferences and song characteristics. The weight adjustment mechanism enables the network to flexibly integrate information from different expert networks, avoiding information redundancy or information loss, improving the accuracy and effectiveness of feature expression, and then performing distribution probability prediction based on the features output by the feature extraction layer, thereby improving the distribution probability prediction accuracy of the candidate distribution network.

[0151] In some optional implementations, in order to more comprehensively and deeply understand the relationship between the user to be recommended and the paid songs, this embodiment introduces domain statistical features, which are used to represent the statistical description of the interactive behavior based on the domain statistical features. The interactive behavior here refers to the interactive behavior between the user to be recommended and the paid songs recommended in history, that is, obtaining the domain statistical features of the user to be recommended under natural distribution and enhanced distribution; next, Figure 4B The song distribution network input part shown inputs the domain statistical features, the music-related features and the song features of the candidate paid songs into the natural distribution enhancement network at the same time to obtain the corresponding first feature vector.

[0152] Among them, domain statistical features can be understood as: quantitative indicators that describe users' interactive behaviors towards songs, especially paid songs, under different distribution strategies such as natural distribution and enhanced distribution; as well as statistical descriptions of user preferences, song popularity and distribution strategy effects reflected by the interactive behaviors.

[0153] The statistical features in this field can be calculated based on the historical behavior data of users, including but not limited to the interactive behaviors of users such as clicking, playing, collecting, and purchasing songs under different distribution strategies, as well as the average play time, click-through rate, conversion rate and other indicators derived from these behaviors. In addition, the statistical features in this field can also include the statistical features of the songs themselves, such as popularity, style, type, etc., as well as a quantitative description of the impact of distribution strategies on user behavior, such as the relative effects of strategies and the degree of user preference for different strategies.

[0154] Regarding the acquisition of domain statistical features, the user's interactive behavior towards songs under different distribution strategies can be collected from the user's historical recommendation list, and features can be extracted to obtain user behavior information, song statistical information, and distribution strategy information; among which, user behavior information can be obtained from the user's average song playback time, click-through rate, collection or purchase conversion rate, etc. under different distribution strategies; song statistical information can be analyzed and determined based on the popularity of the song, the proportion / frequency of paid song purchases, etc.; distribution strategy information can be determined by calculating the user's preference for different distribution strategies, as well as the user's acceptance of paid content under different distribution strategies. Next, the above user behavior information can be summarized to determine the comprehensive behavior performance of users under different distribution strategies, such as the weighted average of indicators such as average playback time, click-through rate, and conversion rate, to construct comprehensive behavior characteristics; based on the statistical information of songs under different distribution strategies, analyze the distribution of song types, styles, popularity and other characteristics that users are exposed to under different distribution strategies, calculate the standard deviation, entropy and other statistics of the distribution, and construct song distribution characteristics; based on the distribution strategy information, compare the differences in user behavior under different distribution strategies, calculate the relative effects between strategies, such as the improvement rate of enhanced distribution relative to natural distribution, and construct distribution strategy characteristics. Then, the comprehensive behavior characteristics, song distribution characteristics and distribution strategy characteristics can be integrated to obtain the final domain statistical characteristics.

[0155] In this embodiment, by introducing domain statistical features, the dimensions of the input data are enriched, providing richer background information for the feature extraction of the natural expert network, and being able to analyze user behavior patterns from a new perspective, thereby more accurately and comprehensively grasping important information such as the user's preference for different paid songs and potential purchase intention.

[0156] In some optional embodiments, Figure 4A The contribution weight shown in , the above embodiment records that the contribution weight is obtained according to the first eigenvector and the second eigenvector, this embodiment provides two specific methods for determining the contribution weight, including:

[0157] The contribution weight can be determined by obtaining the inner product operation result of the first eigenvector and the second eigenvector, wherein the inner product is a scalar value that can measure the similarity or correlation between two vectors. After obtaining the inner product operation result, a certain mapping or conversion function can be used to convert the inner product value into a specific value of the contribution weight, and the value range of the contribution weight is 0-1.

[0158] Alternatively, similarity calculation can be performed based on the first feature vector and the second feature vector through one or more network layers, and the contribution weight can be determined based on the similarity calculation result; wherein the network layer can be a fully connected layer, a convolutional layer or other type of neural network layer, and the network layer is used to perform a series of transformations and calculations on the input first feature vector and the second feature vector to capture the complex relationship between the input features and the similarity between the two, thereby outputting the contribution weight.

[0159] In some optional embodiments, for Figure 4A The song distribution network shown can be trained by the following pre-training steps b1-b4:

[0160] b1, input the song features of the sample user's historical paid songs and the music-related features corresponding to the sample user into the natural distribution expert network, the enhanced distribution expert network and the scene sharing network respectively, and obtain the corresponding fourth feature vector, fifth feature vector and sixth feature vector;

[0161] like Figure 5A Or as shown in 5B, during the training stage of the song distribution network, the music-related features of the sample user and the song features of the sample user's historical paid songs are input as training samples of the song distribution network, and the song features of the historical paid songs include features representing the distribution method of the historical paid songs.

[0162] The training sample input is simultaneously input into three networks, namely the natural distribution expert network, the enhanced distribution expert network and the scene sharing network, for processing. The natural distribution expert network performs feature extraction on the network input and outputs the fourth feature vector, the enhanced distribution expert network performs feature extraction on the network input and outputs the fifth feature vector, and the scene sharing network performs feature extraction on the network input and outputs the sixth feature vector.

[0163] b2, obtaining a sample weight according to the fourth eigenvector and the fifth eigenvector;

[0164] The sample weight can be realized by the inner product operation of the fourth eigenvector and the second eigenvector or by inputting into the network layer for calculation. The specific calculation process is the same as the calculation principle of the contribution weight in the aforementioned embodiment, and the present invention will not repeat it here.

[0165] When the distribution method of historical paid songs in the training sample input is natural distribution, the sample weight is used to represent the degree of information contribution of the enhanced distribution related features extracted by the enhanced distribution expert network, namely the fifth eigenvector, to the natural distribution scenario, and determines the degree of retention of the fifth eigenvector in the current natural distribution scenario.

[0166] When the distribution method of historical paid songs in the training sample input is enhanced distribution, the sample weight is used to represent the degree of information contribution of the natural distribution related features extracted by the natural distribution expert network, namely the fourth eigenvector, to the enhanced distribution scenario, and determines the degree of retention of the fourth eigenvector in the current enhanced distribution scenario.

[0167] b3, obtaining the predicted distribution probability of the historical paid songs according to the sample weight, and the obtained fourth eigenvector, fifth eigenvector and sixth eigenvector;

[0168] In the case where the historical paid songs belong to samples under natural distribution, the sample weight is used to represent the degree of retention of the fifth feature vector extracted by the enhanced distribution expert network under the current natural distribution scenario, see Figure 5A As shown, the sample weight can be multiplied by the fifth eigenvector to obtain a seventh eigenvector; then, the predicted distribution probability of the historical paid songs can be obtained based on the fourth eigenvector, the sixth eigenvector and the seventh eigenvector.

[0169] In the case where the historical paid songs belong to samples under enhanced distribution, the sample weight is used to represent the degree of retention of the fourth feature vector extracted by the natural distribution expert network in the current enhanced distribution scenario, see Figure 5B As shown, the sample weight can be multiplied by the fourth eigenvector to obtain an eighth eigenvector; then, the predicted distribution probability of the historical paid songs can be obtained based on the fifth eigenvector, the sixth eigenvector and the eighth eigenvector.

[0170] Regarding obtaining the predicted distribution probability of the historical paid songs based on the fourth eigenvector, the sixth eigenvector and the seventh eigenvector, or based on the fifth eigenvector, the sixth eigenvector and the eighth eigenvector, reference may be made to the prediction process of step a4 in the aforementioned embodiment.

[0171] b4, calculating the loss function according to the predicted distribution probability of the historical paid songs and the true labels of the historical paid songs, and reversely adjusting the parameters of the song distribution network to minimize the loss function.

[0172] The loss function is used to measure the difference between the predicted probability of the historical paid songs in the training samples by the song distribution network and the real labels of the training samples. In this embodiment, the loss function can adopt cross entropy loss or square loss or other applicable losses. The calculated loss function value reflects the prediction accuracy of the song distribution network in the current training state. The smaller the value of the loss function, the closer the prediction result of the song distribution network is to the actual situation.

[0173] After calculating the loss function, the parameters of the song distribution network can be adjusted through the back propagation algorithm, that is, the weights and biases of the song distribution network are updated according to the gradient of the loss function. In the process of reverse parameter adjustment, optimization algorithms such as stochastic gradient descent SGD, Adam, etc. can be used to gradually reduce the value of the loss function.

[0174] The above steps b1-b4 are repeated until the value of the loss function reaches an acceptable range or reaches a preset number of iterations. During the training process, the song distribution network can gradually learn how to accurately predict the distribution probability of a song based on the input features.

[0175] In the disclosed embodiment, by inputting the song features and music-related features of the sample users' historical paid songs into the natural distribution expert network, the enhanced distribution expert network and the scenario sharing network respectively, the song distribution network can simultaneously learn the features of both natural distribution and enhanced distribution scenarios, and flexibly integrate information from different sources according to the distribution scenario to which the training samples belong, fully considering the importance of different features in different scenarios, thereby improving the accuracy and credibility of the predicted distribution probability.

[0176] In addition, the song distribution network trained in the above manner can learn the correlation and mutual influence between natural distribution and enhanced distribution. Therefore, when using the above-mentioned trained song distribution network to generate song recommendation results in the enhanced distribution scenario, the network can fully consider the impact of natural distribution on enhanced distribution, and can dynamically adjust the impact of the features extracted under the natural distribution method on the enhanced distribution scenario, thereby recommending songs that are more in line with user needs and preferences and can meet the set business goals.

[0177] In some optional implementations, before the present invention introduces the full song trial link, the enhanced distribution method can recommend various paid songs among the full song free trial, fragment free trial, and no trial supported. After the introduction of the full song trial link, the enhanced distribution method will only recommend paid songs that support full song free trial. Based on the historical paid songs under the enhanced distribution method adopted in the training phase, there may be paid songs in the state of fragment free trial and no trial supported. Since songs that do not support trial naturally have a higher conversion rate, it is not conducive to the distribution of paid songs for full song free trial by the song distribution network during the learning process.

[0178] Based on this, this embodiment proposes a method for correcting the song status during the training process to correct the song status, thereby strengthening the distribution of paid songs that can be listened to for free. Figure 6The training phase shown in the figure includes a song distribution network example with song state correction. The song state deviation learning framework is used to simulate the song state deviation in offline training, while ignoring the impact of song state in the inference phase. The song state deviation learning framework is based on the assumption that the user's conversion probability for a song depends on two factors: one is the user's perception of the song's state, and the other is the probability of the user purchasing the song after learning about it. In offline samples, the two probabilities can be simulated through the network. In the online inference process, since the recommended songs are all paid songs that can be listened to in full, the user cannot perceive the song state during the listening process. Therefore, the song distribution network can uniformly process the song state in the online inference phase.

[0179] Based on the song state deviation learning framework in the training stage, i.e., the correction network shown in the figure, the pre-training step of the song distribution network can also include the following correction training process, namely: for each historical paid song in the pre-training sample, the song state of the historical paid song is obtained; the song state includes at least any one of free trial of the whole song, free trial of a fragment, and unsupported trial; the payment state of the historical paid song is simultaneously input into the correction network to obtain the correction score output by the correction network; the predicted distribution probability of the historical paid song is merged with the correction score to obtain an updated distribution probability.

[0180] That is, during the training process, when the music-related features of the sample user and the song features of the historical paid songs of the sample user are input into the song distribution network, the song status of the historical paid songs is simultaneously input into the correction network. The correction network, through deep learning and complex algorithm processing, can output a correction score, which reflects the deviation effect of the song status of the historical paid songs on the distribution probability of the historical paid songs, and represents the amount of adjustment required for the predicted distribution probability predicted based on the music-related features of the sample user and the song features of the historical paid songs under the song status of the historical paid songs. Through the correction network, the deviation effect of different song statuses on the distribution probability can be simulated and predicted more accurately, thereby providing strong support for subsequent fusion processing.

[0181] Regarding the fusion of the predicted distribution probability and the correction score for historical paid songs, the essence is to use the correction score to correct the predicted distribution probability, which can be achieved through weighted summation, multiplication, deducting the correction score from the predicted distribution probability, etc., and can be calculated according to the actual design of the correction score meaning. For example, in the case where the correction score output by the correction network itself represents the distribution probability adjustment value, the difference between the predicted distribution probability and the distribution probability adjustment value can be calculated as the updated predicted distribution probability; in the case where the correction score output by the correction network itself represents the weight of the influence of the song status on the distribution probability, the correction score can be used to perform a weighted calculation on the predicted distribution probability, and the weighted calculation result is used as the updated predicted distribution probability. While ensuring the accuracy of the distribution probability, the potential impact of the song status on the distribution probability is also fully considered, thereby further improving the rationality and credibility of the recommendation results.

[0182] It should be understood that the correction network helps the song distribution network learn and simulate the biased impact of different song states on the predicted distribution probability during the offline training phase, while in the online reasoning phase, the correction processing based on the training phase has eliminated the impact of the song state on the recommendation results, so the reasoning process does not need to use the correction network.

[0183] That is, the task of the correction network in the offline training stage is to analyze the relationship between the song status of historical paid songs and their predicted distribution probability, and to correct the predicted distribution probability of the song distribution network through the correction score. The corrected predicted distribution probability will be calculated with the true label of the historical paid song through the loss function, and the network parameters of the song distribution network will be adjusted by backpropagation. Therefore, after iterative training, the network parameters of the song distribution network learn the influence of the song status on the predicted distribution probability, and eliminate the influence of the song status on the predicted distribution probability. In the online reasoning stage, since the recommended songs are paid songs that support free listening of the entire song, users can freely understand the content of the song during the listening process, and the predicted distribution probability is calculated based on user characteristics and song characteristics.

[0184] In the disclosed embodiment, by adding the song status correction processing in the pre-training step, not only the recommendation effect and accuracy of the song distribution network are improved, but also the distribution probability can be adjusted more scientifically and reasonably when facing different song states, thereby providing users with better recommendation services.

[0185] Exemplary Devices

[0186] After introducing the method of the exemplary embodiment of the present invention, next, refer to Figure 7 A device according to an exemplary embodiment of the present invention is described.

[0187] The implementation process of the functions and effects of each module in the following device is specifically described in the implementation process of the corresponding steps in the above method, which will not be repeated here. For the device embodiment, since it basically corresponds to the method embodiment, the relevant parts can be referred to the partial description of the method embodiment.

[0188] Figure 7 A device for recommending paid songs according to an embodiment of the present invention is schematically shown, and the device comprises:

[0189] The feature acquisition module 701 is used to acquire music-related features of the user to be recommended under natural distribution and enhanced distribution; the music-related features at least include song preference features and payment behavior features of the user to be recommended for paid songs; the natural distribution is used to recommend songs to the user according to the user's song preference information; the enhanced distribution is used to recommend paid songs to the user according to the set business goals;

[0190] Prediction module 702, used to predict the distribution probability of the candidate paid songs according to the music-related features and song features of multiple candidate paid songs in the resource library; the candidate paid songs support free trial listening of the entire song;

[0191] The enhanced distribution result acquisition module 703 is used to select a set number of candidate paid songs from the multiple candidate paid songs as song recommendation results for enhanced distribution according to the distribution probability of the candidate paid songs.

[0192] In some optional embodiments, the device further comprises:

[0193] Obtain song recommendation results for the user to be recommended under natural distribution;

[0194] The song recommendation results of the enhanced distribution are incorporated into the song recommendation results of the natural distribution to obtain the target recommendation results for the user to be recommended.

[0195] In some optional implementations, the prediction module is specifically used to:

[0196] Inputting the music-related features and the song features of the candidate paid songs into a song distribution network to obtain the distribution probability of the candidate paid songs;

[0197] Among them, the pre-trained samples of the song distribution network include: music-related features of sample users; song features of historical paid songs of sample users under natural distribution and enhanced distribution; and the true labels of the pre-trained samples are used to indicate whether the user purchased the historical paid songs.

[0198] In some optional implementations, the prediction module is specifically used to:

[0199] Input the music-related features and the song features of the candidate paid songs into the natural distribution expert network, the enhanced distribution expert network and the scene sharing network respectively to obtain the corresponding first feature vector, the second feature vector and the third feature vector;

[0200] Obtaining a contribution weight according to the first eigenvector and the second eigenvector;

[0201] Multiplying the contribution weight by the first eigenvector to obtain a weighted first eigenvector;

[0202] Obtaining the distribution probability of the candidate paid song according to the second feature vector, the third feature vector and the weighted first feature vector;

[0203] Among them, the natural distribution expert network is used to extract features related to natural distribution; the enhanced distribution expert network is used to extract features related to enhanced distribution; and the scene sharing network is used to extract features common to both natural distribution and enhanced distribution.

[0204] In some optional embodiments, the device further comprises:

[0205] Obtaining the domain statistical features of the user to be recommended under natural distribution and enhanced distribution; the domain statistical features are used to represent the statistical description of the interactive behavior, and the interactive behavior refers to the interactive behavior between the user to be recommended and the paid songs recommended in the past;

[0206] The domain statistical features, the music-related features and the song features of the candidate paid songs are simultaneously input into the natural distribution enhancement network to obtain the corresponding first feature vector.

[0207] In some optional implementations, the prediction module, when used to obtain a contribution weight according to the first feature vector and the second feature vector, includes:

[0208] Obtaining an inner product operation result of the first eigenvector and the second eigenvector; determining the contribution weight according to the inner product operation result;

[0209] Alternatively, similarity calculation is performed through a network layer based on the first feature vector and the second feature vector, and the contribution weight is determined according to the similarity calculation result.

[0210] In some optional implementations, the apparatus further includes a pre-training step of the song distribution network:

[0211] A sample input module, used to input the song features and corresponding music-related features of the historical paid songs into the natural distribution expert network, the enhanced distribution expert network and the scene sharing network, respectively, to obtain the corresponding fourth feature vector, fifth feature vector and sixth feature vector;

[0212] A weight determination module, used for obtaining a sample weight according to the fourth eigenvector and the fifth eigenvector;

[0213] A sample prediction module, used to obtain the predicted distribution probability of the historical paid songs according to the sample weight, and the obtained fourth eigenvector, fifth eigenvector and sixth eigenvector;

[0214] The loss function calculation module is used to calculate the loss function according to the predicted distribution probability of the historical paid songs and the real labels of the historical paid songs, and reversely adjust the parameters of the song distribution network to minimize the loss function.

[0215] In some optional implementations, the sample prediction module is specifically used to:

[0216] In the case where the historical paid songs belong to samples under natural distribution, the sample weight is multiplied by the fifth eigenvector to obtain a seventh eigenvector;

[0217] The predicted distribution probability of the historical paid songs is obtained according to the fourth eigenvector, the sixth eigenvector and the seventh eigenvector.

[0218] In some optional implementations, the sample prediction module is specifically used to:

[0219] In the case where the historical paid song belongs to a sample under enhanced distribution, multiplying the sample weight by the fourth feature vector to obtain an eighth feature vector;

[0220] The predicted distribution probability of the historical paid songs is obtained according to the fifth eigenvector, the sixth eigenvector and the eighth eigenvector.

[0221] In some optional implementations, the pre-training step further includes:

[0222] For each historical paid song in the pre-trained sample, obtain the song status of the historical paid song; the song status includes at least one of free trial listening of the whole song, free trial listening of a fragment, and trial listening not supported;

[0223] Inputting the payment status of the historical paid songs into the correction network at the same time, and obtaining the correction score output by the correction network;

[0224] The predicted distribution probability of the historical paid songs is fused with the correction score to obtain an updated distribution probability.

[0225] Exemplary Media

[0226] After introducing the method and apparatus of the exemplary embodiment of the present invention, next, reference is made to Figure 8 A readable storage medium according to an exemplary embodiment of the present invention is described.

[0227] In this exemplary embodiment, the above method can be implemented by a program product, such as a portable compact disk read-only memory (CD-ROM) and including program code, and can be run on a device, such as a personal computer. However, the program product of the present invention is not limited to this. In this document, a readable storage medium can be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, an apparatus or a device.

[0228] The program product may use any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0229] Computer readable signal media may include data signals propagated in baseband or as part of a carrier wave, in which readable program code is carried. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. Readable signal media may also be any readable medium other than a readable storage medium, which may send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, apparatus, or device.

[0230] The program code contained on the readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wired, optical cable, RE, etc., or any suitable combination of the foregoing.

[0231] Program code for performing the operations of the present invention may be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java, C++, etc., and conventional procedural programming languages ​​such as C or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user computing device, partially on a remote computing device, or entirely on a remote computing device or server. In the case of a remote computing device, the remote computing device may be connected to the user computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computing device (e.g., using an Internet service provider to connect through the Internet).

[0232] Exemplary Computing Devices

[0233] After introducing the method, apparatus and medium of the exemplary embodiments of the present invention, next, reference is made to Fig. 9 A computing device according to an exemplary embodiment of the present invention is described.

[0234] Fig. 9 The computing device 900 shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present invention.

[0235] like Fig. 9 As shown, the computing device 900 is in the form of a general computing device. The components of the computing device 900 may include but are not limited to: at least one processing unit 901, at least one storage unit 902, and a bus 903 connecting different system components (including the processing unit 901 and the storage unit 902).

[0236] The bus 903 includes a data bus, a control bus, and an address bus.

[0237] The storage unit 902 may include a readable medium in the form of a volatile memory, such as a random access memory (RAM) 9021 and / or a cache memory 9022 , and may further include a readable medium in the form of a non-volatile memory, such as a read-only memory (ROM) 9023 .

[0238] The storage unit 902 may also include a program / utility 9025 having a set (at least one) of program modules 9024, such program modules 9024 including but not limited to: an operating system, one or more application programs, other program modules, and program data, each of which or some combination may include an implementation of a network environment.

[0239] The computing device 900 may also communicate with one or more external devices 904 (eg, keyboards, pointing devices, etc.).

[0240] Such communication may be performed via input / output (I / O) interface 905. Furthermore, computing device 900 may also communicate with one or more networks (e.g., a local area network (LAN), a wide area network (WAN), and / or a public network, such as the Internet) via network adapter 906. Fig. 9 As shown, the network adapter 906 communicates with other modules of the computing device 900 via the bus 903. It should be understood that, although not shown in the figure, other hardware and / or software modules can be used in conjunction with the computing device 900, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0241] It should be noted that although several units / modules or sub-units / modules of the paid song recommendation device are mentioned in the above detailed description, this division is only exemplary and not mandatory. In fact, according to an embodiment of the present invention, the features and functions of two or more units / modules described above can be embodied in one unit / module. Conversely, the features and functions of one unit / module described above can be further divided into multiple units / modules to be embodied.

[0242] In addition, although the operations of the method of the present invention are described in a specific order in the drawings, this does not require or imply that the operations must be performed in this specific order, or that all the operations shown must be performed to achieve the desired results. Additionally or alternatively, some steps may be omitted, multiple steps may be combined into one step, and / or one step may be decomposed into multiple steps.

[0243] Although the spirit and principle of the present invention have been described with reference to several specific embodiments, it should be understood that the present invention is not limited to the disclosed specific embodiments, and the division of various aspects does not mean that the features in these aspects cannot be combined to benefit, and such division is only for the convenience of expression. The present invention is intended to cover various modifications and equivalent arrangements included in the spirit and scope of the attached claims.

Claims

1. A method for recommending paid songs, characterized in that: The method comprises: Acquire music-related features of the user to be recommended under natural distribution and enhanced distribution; the music-related features at least include song preference features and payment behavior features of the user to be recommended for paid songs; the natural distribution is used to recommend songs to the user according to the user's song preference information; the enhanced distribution is used to recommend paid songs to the user according to the set business goals; According to the music-related features and the song features of multiple candidate paid songs in the resource library, the distribution probability of the candidate paid songs is predicted; the candidate paid songs support free trial listening of the entire song; According to the distribution probability of the candidate paid songs, a set number of candidate paid songs are selected from the plurality of candidate paid songs as song recommendation results for enhanced distribution.

2. The method according to claim 1, characterized in that The method further comprises: Obtain song recommendation results for the user to be recommended under natural distribution; The song recommendation results of the enhanced distribution are incorporated into the song recommendation results of the natural distribution to obtain the target recommendation results for the user to be recommended.

3. The method according to claim 1, characterized in that The predicting the distribution probability of the candidate paid songs according to the music-related features and the song features of the plurality of candidate paid songs in the resource library includes: Inputting the music-related features and the song features of the candidate paid songs into a song distribution network to obtain the distribution probability of the candidate paid songs; The pre-trained samples of the song distribution network include: music-related features of sample users; song features of historical paid songs of sample users under natural distribution and enhanced distribution; The true label of the pre-trained sample is used to indicate whether the user purchased the historical paid song.

4. The method according to claim 3, characterized in that The step of inputting the music-related features and the song features of the candidate paid songs into a song distribution network to obtain the distribution probability of the candidate paid songs includes: Input the music-related features and the song features of the candidate paid songs into the natural distribution expert network, the enhanced distribution expert network and the scene sharing network respectively to obtain the corresponding first feature vector, the second feature vector and the third feature vector; Obtaining a contribution weight according to the first eigenvector and the second eigenvector; Multiplying the contribution weight by the first eigenvector to obtain a weighted first eigenvector; Obtaining the distribution probability of the candidate paid song according to the second feature vector, the third feature vector and the weighted first feature vector; Among them, the natural distribution expert network is used to extract features related to natural distribution; the enhanced distribution expert network is used to extract features related to enhanced distribution; and the scene sharing network is used to extract features common to both natural distribution and enhanced distribution.

5. The method according to claim 4, characterized in that The method further comprises: Obtaining the domain statistical features of the user to be recommended under natural distribution and enhanced distribution; the domain statistical features are used to represent the statistical description of the interactive behavior, and the interactive behavior refers to the interactive behavior between the user to be recommended and the paid songs recommended in the past; The domain statistical features, the music-related features and the song features of the candidate paid songs are simultaneously input into the natural distribution enhancement network to obtain the corresponding first feature vector.

6. The method according to claim 3, characterized in that The method also includes a pre-training step of the song distribution network: Input the song features of the historical paid songs and the corresponding music-related features into the natural distribution expert network, the enhanced distribution expert network and the scene sharing network respectively to obtain the corresponding fourth feature vector, fifth feature vector and sixth feature vector; Obtaining a sample weight according to the fourth eigenvector and the fifth eigenvector; Obtaining the predicted distribution probability of the historical paid songs according to the sample weight, and the obtained fourth eigenvector, fifth eigenvector, and sixth eigenvector; A loss function is calculated based on the predicted distribution probability of the historical paid songs and the true labels of the historical paid songs, and the parameters of the song distribution network are adjusted inversely to minimize the loss function.

7. The method according to claim 6, characterized in that The pre-training step further comprises: For each historical paid song in the pre-trained sample, obtain the song status of the historical paid song; the song status includes at least one of free trial listening of the whole song, free trial listening of a fragment, and trial listening not supported; Inputting the payment status of the historical paid songs into the correction network at the same time, and obtaining the correction score output by the correction network; The predicted distribution probability of the historical paid songs is fused with the correction score to obtain an updated distribution probability.

8. A device for recommending paid songs, characterized in that: The device comprises: A feature acquisition module is used to acquire music-related features of the user to be recommended under natural distribution and enhanced distribution; the music-related features at least include song preference features and payment behavior features of the user to be recommended for paid songs; the natural distribution is used to recommend songs to the user according to the user's song preference information; the enhanced distribution is used to recommend paid songs to the user according to the set business goals; A prediction module, used to predict the distribution probability of the candidate paid songs according to the music-related features and the song features of multiple candidate paid songs in the resource library; the candidate paid songs support free trial listening of the entire song; The enhanced distribution result acquisition module is used to select a set number of candidate paid songs from the multiple candidate paid songs as song recommendation results for enhanced distribution according to the distribution probability of the candidate paid songs.

9. A readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

10. A computing device, characterized in that: include: Processor, memory; The memory is used to store computer programs; The processor is configured to execute the method according to any one of claims 1 to 7 by calling the computer program.

Images