Preference network model training method and device, preference network model recommendation method and device, equipment and medium

Abstract

The invention discloses a preference network model training method and device, a preference network model recommendation method and device, equipment and a medium, and belongs to the technical field of artificial intelligence. Comprising the steps of obtaining a singer comprehensive feature of each singer based on a singer audio feature of a representative song of each singer in a plurality of singers and a singer meta-information feature of each singer; constructing a plurality of positive sample pairs and a plurality of negative sample pairs based on the similarity between the singer comprehensive features of different singers; the positive sample pair and the negative sample pair are input into a preference network model to be trained, so that the preference network model outputs singer preference characteristics of each singer, and the spatial dimension of the singer comprehensive characteristics is higher than that of the singer preference characteristics; and training the preference network model by using the similarity between the singer preference features of different singers to obtain a trained preference network model. By combining the audio information and the meta-information, personalized recommendation service can be provided for the recommendation object.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of artificial intelligence technology, and in particular to a training method, recommendation method, device, equipment and medium for a preference network model. Background Art

[0002] In the music recommendation system, accurately recommending songs that users are interested in is crucial to improving user experience and the activity of the music platform.

[0003] In the related art, music recommendation systems usually perform similarity recall of singers based on the behavior information of listeners. For example, based on the listeners' playback, collection, rating, sharing and other behavior information, the music recommendation system finds other singers with similar preferences to the listeners from the music library, and then recommends singers that the listeners may like.

[0004] However, the related technology performs better in similarity recall for popular singers, but lacks sufficient behavioral information for unpopular singers, resulting in poor recall performance. Summary of the invention

[0005] The present application provides a training method, recommendation method, device, equipment and medium for a preference network model, and the technical solution is as follows:

[0006] According to one aspect of the present application, a method for training a preference network model is provided, the method comprising:

[0007] Based on the singer audio features of the representative songs of each singer among multiple singers and the singer meta-information features of each singer, the singer comprehensive features of each singer are obtained;

[0008] Based on the similarity between the comprehensive features of different singers, multiple positive sample pairs and multiple negative sample pairs are constructed; wherein the positive sample pair includes the comprehensive features of two singers whose similarity meets a preset similarity condition, and the negative sample pair includes the comprehensive features of two singers whose difference meets a preset difference condition;

[0009] Input the positive sample pairs and the negative sample pairs into the preference network model to be trained, so that the preference network model outputs the singer preference features of each singer; the spatial dimension of the singer's comprehensive features is higher than the spatial dimension of the singer's preference features;

[0010] The preference network model is trained using the similarity between the singer preference features of different singers to obtain a trained preference network model.

[0011] According to one aspect of the present application, a method for recommending a preference network model is provided, the method comprising:

[0012] Obtaining the preferred singers of the recommended object and obtaining the preferred feature set, wherein the preferred feature set includes a set constructed by the singer preference features of each singer output by the trained preference network model, and the trained preference network is obtained by the training method of the preference network model;

[0013] Querying the preference feature set for a first singer preference feature corresponding to the preferred singer;

[0014] Determine a second singer preference feature in the preference feature set whose similarity to the first singer preference feature meets a preset condition, and recommend the singer corresponding to the second singer preference feature as a target singer to the recommendation object.

[0015] According to another aspect of the present application, a computer device is provided, comprising: a processor and a memory, wherein at least one computer program is stored in the memory, and the at least one computer program is loaded and executed by the processor to implement the training method of the preference network model and the recommendation method of the preference network model as described above.

[0016] According to another aspect of the present application, a computer storage medium is provided, in which at least one computer program is stored. The at least one computer program is loaded and executed by a processor to implement the training method of the preference network model and the recommendation method of the preference network model as described above.

[0017] According to another aspect of the present application, a computer program product is provided, which includes a computer program stored in a computer-readable storage medium; the computer program is read and executed from the computer-readable storage medium by a processor of a computer device, so that the computer device executes the training method of the preference network model and the recommendation method of the preference network model as described above.

[0018] The beneficial effects of the technical solution provided by this application include at least:

[0019] In the related art, the music recommendation system usually recommends singers based on the behavioral information of the audience. Due to the lack of behavioral information of unpopular singers, the music recommendation system performs poorly in recommending unpopular singers. This application obtains a large number of representative songs of singers, extracts audio information from the representative songs of the singers, extracts meta information from the singers' own attributes or representative songs, and organically combines the audio information and meta information corresponding to each singer to obtain a comprehensive representation of the singer's information, that is, the singer's comprehensive features; based on the similarity between the singer's comprehensive features of different singers, multiple positive sample pairs and multiple negative sample pairs are constructed, and then the corresponding singer comprehensive features in the multiple positive sample pairs and multiple negative sample pairs are input into the preference network model, so that the preference network model can output singer preference features with similar relationships in the feature vector space. The singer preference features can be used as the basis for the music recommendation system to recommend target singers (including popular singers and unpopular singers) to the recommended objects. This method significantly improves the accuracy of recalling unpopular singers through the complementary advantages between audio information and meta information, thereby providing accurate and personalized recommendation services for the recommended objects. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0021] Figure 1 is a schematic diagram of the architecture of a computer system provided by an exemplary embodiment of the present application;

[0022] Figure 2 is a schematic diagram of a training method for a preference network model provided by an exemplary embodiment of the present application;

[0023] Figure 3 is a flow chart of a method for training a preference network model provided by an exemplary embodiment of the present application;

[0024] Figure 4 is a flow chart of a method for training a preference network model provided by an exemplary embodiment of the present application;

[0025] Figure 5 is a flow chart of a method for training a preference network model provided by an exemplary embodiment of the present application;

[0026] Figure 6 is a schematic diagram of a method for recommending a preference network model provided by an exemplary embodiment of the present application;

[0027] Figure 7is a structural block diagram of a training device for a preference network model provided by an exemplary embodiment of the present application;

[0028] Figure 8 is a structural block diagram of a device for recommending a preference network model provided by an exemplary embodiment of the present application;

[0029] Fig. 9 It is a structural diagram of a computer device provided by an exemplary embodiment of the present application. DETAILED DESCRIPTION

[0030] In order to make the objectives, technical solutions and advantages of the present application clearer, the implementation methods of the present application will be further described in detail below with reference to the accompanying drawings.

[0031] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application. Instead, they are merely examples of devices and methods consistent with some aspects of the present application as detailed in the appended claims.

[0032] The terms used in this application are for the purpose of describing specific embodiments only and are not intended to limit this application. The singular forms of "a", "said" and "the" used in this application and the appended claims are also intended to include plural forms unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items.

[0033] It should be understood that although the terms first, second, etc. may be used in the present disclosure to describe various information, such information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of the present disclosure, the first parameter may also be referred to as the second parameter, and similarly, the second parameter may also be referred to as the first parameter. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0034] It should be noted that the data involved in this application (including but not limited to data used for analysis, training data, recommended data, etc.) are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of relevant countries and regions. For example, the information such as setting operations involved in this application is obtained with full authorization.

[0035] Before introducing the technical solution of the present application, some terms involved in the present application are explained. The following related explanations can be combined arbitrarily with the technical solution of the embodiment of the present application as optional solutions, and they all belong to the protection scope of the embodiment of the present application. The embodiment of the present application includes at least part of the following contents.

[0036] Behavior information: usually refers to the interaction data between users and content, which can be used to analyze user preferences and interests. For example, in a music recommendation system, behavior information can be user click behavior (which singers or songs the user clicked), play behavior (the songs or albums the user actually played), rating behavior (the user's rating of songs), collection behavior (the user's adding certain songs or singers to personal collections), sharing behavior (the user's sharing of songs or singers), etc.

[0037] In the embodiment of the present application, behavior information may also be referred to as operation behavior information.

[0038] Graph Sample and Aggregation (GraphSAGE): is an algorithm for graph neural network models that is designed to process large-scale graph data and is used to learn node representations from graph structured data. Graph neural network models can convert graph data into vector representations (also called embeddings, feature vectors, representation vectors, etc.), and convert nodes in the graph or the entire graph into computable vectors, providing a suitable computing interface for machine learning models to process high-dimensional and complex data forms such as graphs. When mapping nodes to vector representations, as much topological information of the graph as possible is retained, such as making the vector representations of nodes that are close in the graph structure as close as possible in the feature vector space.

[0039] In the embodiment of the present application, the nodes in GraphSAGE represent different singers, the edges in GraphSAGE represent the similarities between different singers, and the weight of the edge represents the degree of similarity between singers. Optionally, the greater the weight of the edge, the greater the similarity between singers, that is, the more similar the singers are; the smaller the weight of the edge, the smaller the similarity between singers, that is, the less similar the singers are. For example, in GraphSAGE, including node 1 (representing singer 1), node 2 (representing singer 2), node 3 (representing singer 3) and node 4 (representing singer 4), the weight of the edge between node 1 and node 2 is 0.2, the weight of the edge between node 1 and node 3 is 0.5, and the weight of the edge between node 1 and node 4 is 0.3, that is, singer 1 and singer 3 are the most similar, and singer 1 and singer 2 are the least similar. Optionally, the similarity relationship between singers is obtained based on the user's behavior information, for example, the similarity relationship between singers is obtained based on the behavior of a large number of users playing, liking, forwarding, commenting on the singer's songs, etc.

[0040] Figure 1 The schematic diagram of the architecture of a computer system provided by an embodiment of the present application is shown. The computer system may include: a terminal 100 and a server 200.

[0041] The terminal 100 may be an electronic device such as a mobile phone, a tablet computer, a vehicle-mounted terminal (vehicle computer), a wearable device, a personal computer (PC), a vehicle-mounted terminal, etc. The terminal 100 may be installed with a client that runs a target application, and the target application may be an application that supports training or recommendation of a preferred network model, which is not limited in this application. In addition, this application does not limit the form of the target application, including but not limited to an application (Application, App), a mini-program, a client, etc. installed in the terminal 100, and may also be in the form of a web page.

[0042] The server 200 may be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server, cloud database, cloud computing, cloud function, cloud storage, network service, cloud communication, middleware service, domain name service, security service, content delivery network (CDN), and cloud server for basic cloud computing services such as big data and artificial palm image recognition platform. The server 200 may be a background server of the above-mentioned target application, used to provide background services for the client of the target application. In some embodiments, the above-mentioned server may also be implemented as a node in a blockchain system.

[0043] The terminal 100 and the server 200 may communicate with each other via a network, such as a wired or wireless network.

[0044] In the training method of the preference network model or the recommendation method of the preference network model provided in the embodiment of the present application, the execution subject of each step may be a computer device, which refers to an electronic device with data calculation, processing and storage capabilities. Figure 1 Taking the computer system shown as an example, the training method of the preference network model or the recommendation method of the preference network model can be executed by the terminal 100, the training method of the preference network model or the recommendation method of the preference network model can be executed by the server 200, or the terminal 100 and the server 200 can interact and cooperate to execute the training method of the preference network model or the recommendation method of the preference network model, and the present application does not limit this.

[0045] For example, Figure 2 A schematic diagram of a training method for a preference network model provided by an exemplary embodiment of the present application is shown. The method is executed by a computer device, which may be Figure 1 The terminal 100 and / or the server 200 are shown.

[0046] In the related art, music recommendation systems usually perform similarity recall of singers based on the behavioral information of listeners. For example, based on the behavioral information of listeners such as playing, collecting, rating, and sharing of songs (behavioral information represents user preferences), music recommendation systems find other singers with similar preferences to the listeners in the music library, and then recommend singers that the user may like.

[0047] Usually, a graph neural network model is deployed in a music recommendation system. The graph neural network model is a network model trained based on the user's behavioral information. The music recommendation system recommends singers that the user may like based on the graph neural network model. However, since the graph neural network model is trained based on the user's behavioral information, the user's interactive behavior data on popular singers is sufficient, and the interactive behavior data on unpopular singers is less. Therefore, when making recommendations, the music recommendation system performs better on popular singers, that is, the similarity recall performance of popular singers is better; while unpopular singers lack sufficient behavioral information, the music recommendation system performs poorly on unpopular singers, that is, the similarity recall performance of unpopular singers is poor.

[0048] In order to solve the above problems, the present application proposes a training method for a preference network model, by selecting a preset number of representative songs of a singer, extracting audio information and the corresponding meta-information of the singer from the preset number of representative songs, and aligning the singer audio features corresponding to the audio information and the singer meta-information features corresponding to the meta-information, splicing the aligned audio features and the aligned meta-information features, and obtaining the singer comprehensive features, which are used to characterize the singer's comprehensive information features. Multiple singer comprehensive features are input into the preference network model, and the preference network model learns the similarity relationship between different singer comprehensive features in the process of reducing the dimension of the singer comprehensive features, and outputs the singer preference features with similar relationships in the feature vector space. The singer preference features corresponding to singers with similar styles are close in distance in the feature vector space, and the singer preference features corresponding to singers with different styles are far away in the feature vector space. In the subsequent recommendation process, according to the preferred singers of the recommended object, the target singers corresponding to the singer preference features with similar distances in the preference feature set can be recommended to the recommended object.

[0049] Combined with reference Figure 2In the music recommendation system, a first number of singers are selected from a music database, and each singer in the first number of singers corresponds to a second number of representative songs 10. The representative works in the representative songs 10 are the first n works of each singer sorted from high to low according to the number of plays, and n is a positive integer. Optionally, the representative songs 10 corresponding to each singer are input into an audio feature extraction model 20, and the audio feature extraction model 20 extracts audio information and outputs audio features 30. Specifically, the audio feature extraction model 20 performs audio extraction on each representative work of each singer to obtain audio features 30 corresponding to each representative work. Then, the audio features corresponding to multiple representative works are averaged or weighted to obtain the singer audio features 40, and the singer audio features 40 are used to characterize the comprehensive audio features of the singer. Among them, the audio information includes but is not limited to at least one of the pitch, rhythm, speed, and melody corresponding to the representative songs 10.

[0050] Exemplarily, the representative songs 10 of singer 1 are selected, and the representative songs 10 include six representative works, namely, representative work 1, representative work 2, ..., representative work 6. The audio features 31 corresponding to representative work 1, the audio features 32 corresponding to representative work 2, the audio features 33 corresponding to representative work 3, ... are obtained through the audio feature extraction model 20. The audio features corresponding to the six representative works are averaged or weighted averaged to obtain the singer audio features 40. Among them, the average processing or weighted average processing can help eliminate the randomness of a single representative work in the feature extraction process, thereby obtaining more stable and representative audio features.

[0051] In some embodiments, the meta information 50 is extracted from the album corresponding to the representative song 10. A part of the meta information 50 may be information representing the singer's own attributes, such as at least one of gender, genre, and country; another part of the meta information 50 may be information about the representative song, such as the language of the representative song and the year of release of the representative song. Optionally, the meta information 50 is digitized, and each category of the meta information 50 (such as gender, genre, country, and language, etc.) is converted into a digital form to obtain the singer meta information feature 60. The singer meta information feature 60 is a multidimensional label vector, and each label represents a category of meta information.

[0052] In some embodiments, the singer audio features 40 and the singer meta information features 60 are in different spatial dimensions (i.e., the dimensions of the feature vector space). Optionally, the singer audio features 40 and the singer meta information features 60 are input into the alignment model 70, and the singer audio features 40 and the singer meta information features 60 are aligned through the alignment model 70. Optionally, the alignment model 70 includes an audio alignment model 71 and a meta information alignment model 72, the singer audio features 40 are input into the audio alignment model 71, and the singer meta information features 60 are input into the meta information alignment model 72, so as to obtain the singer audio features 40 and the singer meta information features 60 in the same spatial dimension. In the alignment model 70, the two vectors of the same singer constitute a positive pair, that is, the singer audio features 40 of the same singer and the singer meta information features 60 corresponding thereto constitute a positive pair; the two vectors of different singers constitute a negative pair, that is, the singer audio features 40 of different singers and the unrelated singer meta information features 60 constitute a negative pair. This makes the singer audio features 40 and the singer meta information features 60 of the same singer closer in the feature vector space, while the features of different singers are more dispersed.

[0053] In some embodiments, the aligned singer audio features 40 and singer meta-information features 60 are concatenated to obtain singer comprehensive features 80 , and the singer comprehensive features 80 are used to characterize the comprehensive information features of the singer.

[0054] Optionally, a plurality of different singer comprehensive features 80 are input into the preference network model 90, and the preference network model 90 outputs singer preference features 91 for characterizing the singer comprehensive features after dimensionality reduction. Optionally, a sample pair set is constructed based on the similarity of different singers, and the sample pair set includes positive sample pairs and negative sample pairs.

[0055] In one implementation, singers in GraphSAGE whose similarity meets a preset similarity condition are selected as positive sample pairs (i.e., the comprehensive characteristics of the two singers whose similarity meets the preset similarity condition), and singers in GraphSAGE whose difference meets a preset difference condition are selected as negative sample pairs (i.e., the comprehensive characteristics of the two singers whose difference meets the preset similarity condition), and the positive sample pairs and the negative sample pairs are input into the preference network model 90 to be trained.

[0056] In some embodiments, the singer comprehensive features 80 of multiple different singers (similar singer comprehensive features are positive sample pairs, and dissimilar singer comprehensive features are negative sample pairs) are input into the preference network model 90 for training. The preference network model 90 can learn the similarity relationship between different singer comprehensive features 80, and output singer preference features 91 having similar relationships in the feature vector space.

[0057] In some embodiments, the preference network model 90 may be optimized. Optionally, the similarity between the singer preference features of different singers is obtained, and the similarity between the attribute information of different singers is obtained; based on the similarity between the singer preference features of different singers and the similarity between the attribute information of different singers, a similarity loss value is determined, and the preference network model is optimized based on the similarity loss value. The similarity between the attribute information of different singers is obtained based on the graph neural network model.

[0058] In some embodiments, the nodes in GraphSAGE represent different singers, the edges in GraphSAGE represent the similarities between different singers, the weights of the edges represent the similarities between singers, and the similarity results of GraphSAGE are the similarities between different singers. Optionally, the greater the weight of the edge, the greater the similarity between singers, that is, the more similar the singers are; the smaller the weight of the edge, the smaller the similarity between singers, that is, the less similar the singers are.

[0059] In some embodiments, a preference feature set is obtained, which is a set constructed by singer preference features corresponding to each singer output by the trained preference network model 90. Optionally, singer preference features 91 corresponding to singers with similar styles (i.e., having a similar relationship, or a high degree of similarity) are close in distance on the feature vector space, and singer preference features 91 corresponding to singers with very different styles (i.e., not having a similar relationship, or a low degree of similarity) are far away in the feature vector space.

[0060] In the subsequent recommendation process, the music recommendation system can query the first singer preference feature corresponding to the preferred singer in the preference feature set based on the preferred singer of the recommendation object, and recommend the singer corresponding to the second singer preference feature as the target singer to the recommendation object by determining the second singer preference feature in the preference feature set whose similarity with the first singer preference feature meets the preset conditions.

[0061] Figure 3 is a flowchart of a method for training a preference network model provided by an exemplary embodiment of the present application. The method can be executed by a computer device, which can be Figure 1 The method comprises the following steps:

[0062] Step 210: obtaining a singer comprehensive feature of each singer based on the singer audio features of the representative songs of each singer among the multiple singers and the singer meta-information features of each singer;

[0063] In some embodiments, representative songs of multiple singers are obtained from a music database, wherein each of the multiple singers has at least one representative song (usually, the number of representative songs is multiple).

[0064] In some embodiments, the representative songs usually represent the music style and popularity of each singer. Optionally, the representative songs are usually the first n songs of the singer selected from high to low according to any one of the following: the number of plays, popularity, number of comments, number of favorites, and number of reposts. Among them, the first n songs can be defined as the representative songs of the singer, and n is a positive integer. The embodiment of the present application does not limit the method for determining the representative songs of each singer.

[0065] In some embodiments, the singer audio features are obtained based on the audio information of the representative songs, and the singer audio features refer to the audio features of the representative songs. Optionally, the audio information includes but is not limited to at least one of the pitch corresponding to the representative songs, the rhythm corresponding to the representative songs, the speed corresponding to the representative songs, and the melody corresponding to the representative songs. The singer audio features include but are not limited to at least one of the pitch features corresponding to the representative songs, the rhythm features corresponding to the representative songs, the speed features corresponding to the representative songs, and the melody features corresponding to the representative songs.

[0066] Among them, pitch refers to the high or low pitch of the sound of the representative song, and the pitch can be represented by musical notes, such as C, D, E and other musical letters represent different pitches. Rhythm refers to the changes and repetitions of the sound beats in the representative songs, such as brisk rhythm, slow rhythm, rapid rhythm, etc. Speed ​​refers to the speed of the performance of the representative song, usually expressed as the number of beats per unit time, for example, the speed can be fast, slow, etc. Melody is the pitch change formed by a series of notes arranged in a specific order, for example, melody can be divided into continuous melody, jumping melody, parallel melody, variation melody, etc.

[0067] In some embodiments, the singer meta-information feature is obtained based on the singer's meta-information, and the singer meta-information feature refers to the meta-information feature corresponding to the singer's meta-information. Part of the meta-information may be information representing the singer's own attributes, such as at least one of gender, genre, and country; another part of the meta-information may be information about the singer's representative songs, such as the language of the representative songs and the year of release of the representative songs. The meta-information does not directly constitute the sound content of the representative songs, but is usually stored in the tags of the representative songs, providing additional information such as the background and attributes of the representative songs.

[0068] Optionally, the meta information includes but is not limited to at least one of the gender corresponding to the singer, the genre corresponding to the singer, the country corresponding to the singer, the language corresponding to the representative song, and the release year of the representative song. Among them, the gender indicates whether the singer is male, female or a group (such as a band). The genre indicates the type or style of music to which the singer belongs, such as rock, jazz, pop, classical, hip-hop, etc. The country indicates the country or region of origin of the singer or band, which can reflect the cultural background of the representative song. The language indicates the language used by the singer when singing the song, such as Chinese, English, Spanish, etc. The release year is the year when the representative song or album was released, which helps to understand the time background of the representative song.

[0069] In some embodiments, when the languages ​​of multiple representative songs are different, the language with the highest frequency of use is selected from the languages ​​of the multiple representative songs as the language of the representative song. For example, among the six representative songs of singer A, the language of four representative songs is Chinese and the language of two representative songs is English, and Chinese can be used as the language of the representative song.

[0070] In some embodiments, when the release years of multiple representative songs are different, a representative release year can be selected from the release years of multiple representative songs. Alternatively, the first-ranked representative song is selected from multiple representative songs, and the release year of the first-ranked representative song is used as the release year of the representative song. For example, among the six representative songs of singer B, the release years are 1999, 2000, 2000, 2000, 2002, and 2003 respectively. 2000 is considered to be a representative release year, and the release year of the representative song is 2000.

[0071] In some embodiments, audio features are extracted from the song audio corresponding to each singer's representative song to obtain singer audio features, and text features are extracted from the meta information corresponding to each singer to obtain singer meta information features. Optionally, the singer meta information features can be regarded as a multi-dimensional label vector, each label representing a category of meta information.

[0072] Among them, the singer's comprehensive features are used to characterize the singer's comprehensive information features.

[0073] The singer audio features are used to characterize the singer's comprehensive audio features, and the singer meta information features are used to characterize the singer's comprehensive text features. Optionally, the singer audio features and the singer meta information features are in different spatial dimensions. For example, the dimension of the singer audio features is

[1024] , and the dimension of the singer meta information features is

[50] . Directly concatenating the singer audio features and the singer meta information features cannot achieve a good concatenation effect. Therefore, the singer audio features and the singer meta information features need to be aligned in spatial dimensions.

[0074] In some embodiments, the singer audio features and the singer meta-information features are aligned from different spatial dimensions to the same spatial dimension. For example, the dimension of the singer audio features is aligned from

[1024] to

[40] , and the singer meta-information features are aligned from

[50] to

[40] . The aligned singer audio features and the aligned singer meta-information features are concatenated to obtain the singer comprehensive features of each singer. The singer comprehensive features, as a comprehensive vector, can comprehensively represent the singer's musical style.

[0075] Step 220: constructing a plurality of positive sample pairs and a plurality of negative sample pairs based on the similarity between the comprehensive features of different singers;

[0076] Among them, the positive sample pair includes the singer comprehensive features of two singers whose similarity meets the preset similarity condition, and the negative sample pair includes the singer comprehensive features of two singers whose difference meets the preset difference condition.

[0077] The preset similarity condition indicates that the similarity of the singer comprehensive features corresponding to the two singers reaches a similarity threshold. The preset difference condition indicates that the difference of the singer comprehensive features corresponding to the two singers reaches a difference threshold. The present application does not limit the specific values ​​of the similarity threshold and the difference threshold.

[0078] In some embodiments, the similarity and difference between the positive sample pair and the negative sample pair are obtained based on the graph neural network model. Optionally, singers with higher confidence are selected from the graph neural network model as training samples to train the preference network model. Among them, singers with higher confidence represent singers with more sufficient operation behavior information, for example, singers with higher confidence are singers with high playback volume, and for another example, singers with higher confidence are singers with high collection volume.

[0079] Exemplarily, the graph neural network model is GraphSAGE. The nodes in GraphSAGE represent different singers. For example, in GraphSAGE, the nodes corresponding to singer 1 and singer 2 are the closest (satisfying the preset similarity condition), and the nodes corresponding to singer 1 and singer 5 are the farthest (satisfying the preset difference condition). Singer 1 and singer 2 can be used as a positive sample pair, that is, the singer comprehensive features corresponding to singer 1 and the singer comprehensive features corresponding to singer 2 are used as a positive sample pair; singer 1 and singer 5 can be used as a negative sample pair, that is, the singer comprehensive features corresponding to singer 1 and the singer comprehensive features corresponding to singer 5 are used as a negative sample pair.

[0080] Step 230: inputting the positive sample pairs and the negative sample pairs into the preference network model to be trained, so that the preference network model outputs the singer preference features of each singer;

[0081] Among them, the preference network model is used to output the singer preference characteristics of each singer based on the comprehensive singer characteristics of each singer.

[0082] In some embodiments, the positive sample pairs and the negative sample pairs are respectively input into the preference network model to be trained, that is, the singer pairs with close distance and the singer pairs with distant distance in the graph neural network model (such as GraphSAGE) are respectively input into the preference network model, that is, the singer comprehensive feature pairs corresponding to the singer pairs with close distance in the graph neural network model (such as GraphSAGE) (the singer comprehensive features of two singers whose similarity meets the preset similarity condition) and the singer comprehensive feature pairs corresponding to the singer pairs with distant distance (the singer comprehensive features of two singers whose difference meets the preset difference condition) are respectively input into the preference network model, so that the preference network model outputs the singer preference features of each singer.

[0083] Optionally, the spatial dimension of the singer's comprehensive feature is higher than the spatial dimension of the singer's preference feature. The singer's preference feature is a singer's comprehensive feature after dimensionality reduction, and the singer's preference feature of each singer can be used to characterize the music style of each singer.

[0084] Among them, the singer comprehensive features are used to represent the comprehensive information of the corresponding singer. The similarity between different singers is learned by training the preference network model, and the singer preference features with similarity are output based on the singer comprehensive features. It can also be understood that the singer comprehensive features in the preference network model learn the similarity between different singers in the process of dimensionality reduction, so that the singer preference features output by the preference network model have similarity.

[0085] Step 240: train the preference network model using the similarity between the singer preference features of different singers to obtain a trained preference network model.

[0086] In some embodiments, the preference network model outputs singer preference features of different singers, and the similarity between the singer preference features of different singers is obtained based on the singer preference features of different singers. Optionally, the preference network model is trained using the similarity between the singer preference features of different singers, and the similarity between the singer preference features is characterized based on the distance between the singer preference features of different singers.

[0087] In some embodiments, the preference network model is trained based on a triplet loss function. Optionally, the preference network model is trained with the goal of minimizing the first distance between the anchor sample and the positive sample and maximizing the second distance between the anchor sample and the negative sample to obtain a trained preference network model. The trained preference network model can output singer preference features that have similar relationships in the feature vector space.

[0088] Optionally, by obtaining the singer preference features, a preference feature set is constructed based on the singer preference features. The preference feature set is a set constructed from the singer preference features corresponding to each singer output by the trained preference network model, and the preference feature set includes multiple singer preference features. In the subsequent recommendation process, singers corresponding to singer preference features with similar distances in the preference feature set can be recommended to the recommended object based on the preferred singers of the recommended object. Specifically, the singer preference features corresponding to the preferred singers are queried in the preference feature set, and the target singer preference features with similar distances to the singer preference features corresponding to the preferred singers in the preference feature set are determined, and the target singer corresponding to the target singer preference features is recommended to the recommended object. Based on this method, the music recommendation system can quickly query popular singers and unpopular singers with similar preferences to the recommended object.

[0089] Based on the audio information and meta information, the singer audio features and singer meta information features are obtained respectively.

[0090] In some embodiments, the audio information corresponding to each singer's representative songs is extracted through an audio feature extraction model to obtain the singer's audio features, and the singer's meta-information features are obtained by performing digital processing on the meta-information corresponding to each singer.

[0091] ·Targeted at singer audio features

[0092] Optionally, the audio features of at least one representative song of each singer are integrated to obtain the singer audio features of each singer.

[0093] In some embodiments, the audio features of at least one representative song of each singer are integrated to obtain the singer audio features of each singer. Optionally, the audio information corresponding to at least one representative song of each singer is extracted through an audio feature extraction model to obtain multiple audio features, and the multiple audio features are averaged or weighted to obtain the singer audio features of each singer.

[0094] In some embodiments, each singer corresponds to at least one (usually multiple) representative songs. The audio features can be extracted from the audio information corresponding to each representative song through the audio feature extraction model. By extracting the corresponding multiple audio features from the audio information corresponding to multiple representative songs, the multiple audio features are averaged or weighted to obtain the singer audio features corresponding to the singer.

[0095] In some embodiments, the audio feature extraction model is used to extract audio features from an audio signal representing a song. For example, the audio feature extraction model is a MERT model or a MusicFM model. Optionally, the audio features extracted by the audio feature extraction model are a two-dimensional matrix [T, N]. Wherein T represents the number of time frames, and N represents the feature dimension of the audio feature in each time frame, for example, N=1024, indicating that each time frame has 1024 feature values. Optionally, for the feature vector of each time frame, the average is performed along the time dimension (T), that is, the values ​​of each feature dimension of all time frames are averaged to reduce the changes in the time dimension, so that a more stable feature representation can be obtained, that is, the audio feature is obtained.

[0096] In some embodiments, each singer has at least one representative song. In the embodiments of the present application, the representative songs of each singer are mainly used as an example for explanation, which will not be repeated below. Optionally, each representative song can obtain time-averaged audio features through feature extraction by an audio feature extraction model, and multiple representative songs can obtain corresponding multiple audio features. Average processing or weighted processing is performed on the multiple audio features, that is, average processing or weighted processing is performed on the multiple audio features corresponding to the representative songs of each singer, and a singer audio feature that integrates multiple audio features can be obtained.

[0097] In some embodiments, the singer audio feature is obtained by averaging multiple audio features. Averaging is a simple feature fusion method, that is, multiple audio feature vectors are added together and then divided by the number of representative songs to obtain the final singer audio feature. Exemplarily, singer A includes three representative songs, where representative work 1 corresponds to audio feature 31, representative work 2 corresponds to audio feature 32, and representative work 3 corresponds to audio feature 33. The audio features corresponding to the three representative works are averaged to obtain a final singer audio feature. The dimension of the singer audio feature is the same as the dimension of the audio feature. For example, the dimension of audio feature 31, audio feature 32, and audio feature 33 is N, and the dimension of the singer audio feature obtained after averaging is also N.

[0098] In some embodiments, weighted processing is performed on multiple audio features to obtain the singer's audio features. Among them, weighted processing refers to weighted averaging of the multiple extracted audio features, and assigning different weights to different audio features. Each audio feature is multiplied by its corresponding weight coefficient, and then these weighted audio features are added up to obtain the final singer's audio features. Optionally, the weight coefficients of multiple audio features are determined based on the behavioral data corresponding to each representative song. For example, the weight coefficients of the corresponding audio features are determined based on the corresponding playback volume, collection volume or comment volume of each representative song, and this application does not limit this.

[0099] Exemplarily, based on the playback volume of masterpieces 1, 2, and 3, the weight coefficient of audio feature 31 is determined to be 0.3, the weight coefficient of audio feature 32 is 0.3, and the weight coefficient of audio feature 33 is 0.4. Audio feature 31, audio feature 32, and audio feature 33 can be multiplied by their respective weight coefficients and added together to obtain the final singer's audio features.

[0100] In an embodiment of the present application, an audio feature extraction model is used to analyze multiple representative songs of a singer to obtain multiple audio features of the singer. The multiple audio features are averaged or weighted to obtain the singer's audio features. This can reduce the randomness or noise impact of individual audio features, making the final singer's audio features more stable and reliable, and helping to better describe the singer's overall performance.

[0101] · Targeting singer meta information features

[0102] Optionally, digital processing is performed on the meta information of each singer to obtain singer meta information features of each singer.

[0103] In some embodiments, the singer meta-information features are obtained by performing digital processing on the meta-information corresponding to each singer. The digital processing is used to convert the meta-information into numerical data. The singer meta-information features can be considered as a multi-dimensional label vector, and each dimensional label represents a category of meta-information.

[0104] In some embodiments, the category of each meta-information included in the meta-information corresponding to the singer is first confirmed, and then the meta-information is digitized according to the category. The digitization is used to convert the meta-information into numerical data. Optionally, the digitization includes label encoding or one-hot encoding. Label encoding is a method of converting a categorical variable into numerical data, in which each category value is assigned an integer. One-hot encoding creates a new binary column for each category, indicating whether the category exists.

[0105] Illustratively, the following Table 1 provides an example description of meta information corresponding to a singer.

[0106] Table 1

[0107] Artist ID gender school nation language Year of release Singer A male Rock Country A English 2020 Singer B female Jazz Country B English 2018 Singer C combination Popularity Country C Chinese 2023

[0108] It should be noted that, when there are multiple representative songs of singer A, singer B and singer C in Table 1 above, the language corresponding to the representative song can be considered as the language with the highest frequency of use selected from the languages ​​of the multiple representative songs as the language of the representative song; the release year of the representative song can be considered as a representative release year selected from the release years of the multiple representative songs. This application does not limit this.

[0109] Exemplarily, the meta information is coded by category. For the gender of the singer included in the meta information: male is 0, female is 1, and group (band includes male and female) is 3; for the genre of the singer included in the meta information: rock is 4, jazz is 5, and pop is 6; for the country of the singer: country A is 7, country B is 8, and country C is 9; for the language of the representative song: English is 10, and Chinese is 11; for the release year of the representative song: if the release year is before 2019, it corresponds to 12, and if the release year is after 2019, it corresponds to 13. The meta information of singers A, B and C is label encoded according to type to obtain the singer meta information features. The singer meta information features corresponding to singer A can be expressed as [0, 4, 7, 10, 13], the singer meta information features corresponding to singer B can be expressed as [1, 5, 8, 10, 12], and the singer meta information features corresponding to singer C can be expressed as [3, 6, 9, 11, 13].

[0110] Exemplarily, the meta-information is one-hot encoded according to the category. In one-hot encoding, each category is converted into a binary vector. If the category exists, the corresponding position is 1; otherwise, it is 0. Among them, the binary vector corresponding to the gender is [1, 0, 0] when the gender is male, the binary vector corresponding to the gender is [0, 1, 0] when the gender is female, and the binary vector corresponding to the gender is [0, 0, 1] when the gender is a combination; the genre, country, and language are encoded in the same way as above; for the meta-information including the year of release of the representative song: when the year of release of the representative song is before 2019, the corresponding binary vector is [1, 0], and when the year of release of the representative song is after 2019, the corresponding binary vector is [0, 1]. The meta-information of the above singers A, B and C is one-hot encoded according to the type to obtain the singer meta-information features, where the singer meta-information features corresponding to singer A can be expressed as [1, 0, 0; 1, 0, 0; 1, 0, 0; 1, 0; 0, 1], the singer meta-information features corresponding to singer B can be expressed as [0, 1, 0; 0, 1, 0; 0, 1, 0; 1, 0; 1, 0], and the singer meta-information features corresponding to singer C can be expressed as [0, 0, 1; 0, 0, 1; 0, 0, 1; 0, 1; 0, 1].

[0111] · Comprehensive characteristics of singers

[0112] Optionally, the singer comprehensive features of each singer are obtained based on the singer audio features of each singer and the singer meta-information features of each singer.

[0113] The singer audio features are used to characterize the singer's comprehensive audio features. The singer meta-information features are used to characterize the singer's comprehensive text features. Optionally, for each singer, the singer audio features and singer meta-information features corresponding to each singer are concatenated to obtain the singer comprehensive features of each singer. The singer comprehensive features of each singer are used to characterize the comprehensive information features of each singer.

[0114] In an embodiment of the present application, by digitally processing the metadata of each singer, the singer meta-information features (multi-dimensional label vector) corresponding to each singer are obtained. Through digital processing, different categories of meta-information can be converted into numerical data, which is helpful for unified processing in machine learning and data analysis. By constructing a multi-dimensional label vector, richer information features can be captured and the ability to describe the singer is enhanced.

[0115] Align singer audio features with singer meta information features

[0116] In some embodiments, the singer audio features and the singer meta information features are in different spatial dimensions. Optionally, the singer audio features and the singer meta information features are input into an alignment model, and the singer audio features and the singer meta information features are aligned through the alignment model. Optionally, the alignment model includes an audio alignment model and a meta information alignment model.

[0117] Figure 4 A flowchart of a method for training a preference network model provided by an exemplary embodiment of the present application is shown, wherein the above step 210 can be replaced by step 211, step 212 and step 213.

[0118] Step 211: inputting the audio features of each singer into a pre-trained audio alignment model, and outputting the aligned audio features of the singer;

[0119] Among them, the singer audio features are used to characterize the comprehensive audio features of the singer.

[0120] In some embodiments, the audio alignment model is a machine learning or deep learning model that processes audio data, and its main purpose is to align multiple singer audio features and convert them into a preset dimension. The output of the audio alignment model is an aligned singer audio feature, and the dimension of the aligned singer audio feature is usually much smaller than the dimension of the singer audio feature before alignment, thereby obtaining a more compact singer audio feature.

[0121] In some embodiments, an alignment function is designed in the audio alignment model, and the alignment function is used to form a positive pair of two vectors of the same singer, that is, the singer audio features of the same singer and the corresponding singer meta-information features form a positive pair, and form a negative pair of two vectors of different singers, that is, the singer audio features of different singers and unrelated singer meta-information features form a negative pair. This design makes the singer audio features and singer meta-information features of the same singer closer in the feature vector space, while the singer audio features and singer meta-information features of different singers are more dispersed in the feature vector space.

[0122] Optionally, the alignment function in the audio alignment model is as follows:

[0123]

[0124] Among them, L1 represents the alignment function based on the audio features of the training singer, N represents the number of positive or negative pairs, k and j represent indexes, that is, traversing the positive or negative pairs. represents the audio features of the singer in the kth pair, represents the singer meta-information feature in the kth pair, represents the singer meta-information feature in the jth pair, and Constitute a positive pair, and constitute a negative pair, sim(·,·) represents the similarity function, is the similarity between the singer audio features in the kth pair and the singer meta-information features in the kth pair, usually calculated using cosine similarity. It is the similarity between the singer audio features in the kth pair and the singer meta-information features in the jth pair, which is usually calculated using cosine similarity. τ represents the temperature coefficient, which is used to control the scaling of the similarity and helps to adjust the separation of positive and negative pairs.

[0125] Step 212: inputting the singer meta information features of each singer into the pre-trained meta information alignment model, and outputting the singer meta information features after alignment processing;

[0126] Among them, the singer meta-information features are used to characterize the comprehensive text features of the singer.

[0127] In some embodiments, the meta information alignment model is a machine learning or deep learning model that processes metadata (or understood as text data), and its main purpose is to align multiple singer meta information features and convert them into a preset dimension. The output of the meta information alignment model is an aligned singer meta information feature, and the dimension of the singer meta information feature after alignment is usually smaller than the dimension of the singer meta information feature before alignment, thereby obtaining a more compact singer meta information feature.

[0128] In some embodiments, an alignment function is designed in the meta-information alignment model, and the alignment function is used to form a positive pair of two vectors of the same singer, that is, the singer audio features of the same singer and the corresponding singer meta-information features form a positive pair, and form a negative pair of two vectors of different singers, that is, the singer audio features of different singers and unrelated singer meta-information features form a negative pair. This design makes the singer audio features and singer meta-information features of the same singer closer in the feature vector space, while the singer audio features and singer meta-information features of different singers are more dispersed in the feature vector space.

[0129] Optionally, the alignment function in the meta-information alignment model is as follows:

[0130]

[0131] Among them, L2 represents the alignment function based on the meta-information features of the training singers, N represents the number of positive or negative pairs, k and j represent indexes, that is, traversing the positive or negative pairs. represents the singer meta-information feature in the kth pair, represents the audio features of the singer in the kth pair, represents the singer meta-information feature in the kth pair, represents the audio features of the singer in the jth pair, and Constitute a positive pair, and constitute a negative pair, sim(·,·) represents the similarity function, is the similarity between the singer meta information feature in the kth pair and the singer audio feature in the kth pair, usually calculated using cosine similarity. It is the similarity between the singer meta-information features in the kth pair and the singer audio features in the jth pair, which is usually calculated using cosine similarity. τ represents the temperature coefficient, which is used to control the scaling of the similarity and helps to adjust the separation of positive and negative pairs.

[0132] In some embodiments, the InFoNCE loss function of the alignment model can be expressed as:

[0133]

[0134] Among them, L InFoNCE Represents the loss function of the alignment model, L1 represents the alignment function based on the audio features of the training singer, and L2 represents the alignment function based on the meta-information features of the training singer.

[0135] Step 213: concatenate the singer audio features after the alignment processing and the singer meta-information features after the alignment processing to obtain the singer comprehensive features of each singer.

[0136] In some embodiments, the singer audio features are input into the audio alignment model, and the singer audio features after alignment are output based on the audio alignment model, and the singer meta information features are input into the meta information alignment model, and the singer meta information features after alignment are output based on the meta information alignment model. Optionally, the singer audio features after alignment and the singer meta information features after alignment are features of the same dimension.

[0137] In some embodiments, the aligned singer audio features and the aligned singer meta information features are concatenated to obtain the singer comprehensive features of each singer in a preset dimension, and the singer comprehensive features are used to characterize the singer's comprehensive information features. The dimension of the singer comprehensive feature is the sum of the dimension of the aligned singer audio features and the dimension of the aligned singer meta information features.

[0138] For example, the dimension of the singer's audio features after alignment processed by the audio alignment model output is

[40] , and the dimension of the singer's meta-information features after alignment processed by the meta-information alignment model output is

[40] . By concatenating the singer's audio features after alignment and the singer's meta-information features after alignment, a comprehensive singer feature with a dimension of

[80] can be obtained.

[0139] In an embodiment of the present application, by aligning and splicing the singer's audio features and the singer's meta-information features, the content features of the audio signal and the singer's meta-information can be effectively combined to form a more comprehensive feature representation (singer's comprehensive features), which can better capture the diversity and complexity of the singer.

[0140] Training preference network model based on triplet loss function

[0141] In some embodiments, the preference network model is trained based on a triplet loss function, and the triplet loss function is determined based on positive sample pairs and negative sample pairs, wherein the positive sample pairs are composed of an anchor sample and a positive sample, and the negative sample pairs are composed of an anchor sample and a negative sample.

[0142] Optionally, the preference network model is trained with the goal of minimizing the first distance between the anchor sample and the positive sample and maximizing the second distance between the anchor sample and the negative sample to obtain a trained preference network model. Optionally, the similarity between the singer preference features of different singers is characterized based on the distance. The first distance and the second distance are used to characterize the similarity between the singer preference features of different singers.

[0143] Alternatively, the triplet loss function is expressed as follows:

[0144] L=max(d(a,p)-d(a,n)+margin,0)

[0145] Where L represents the value of the triplet loss function, a represents the anchor sample, p represents the positive sample, n represents the negative sample, d(a,p) represents the distance between the positive sample pair, that is, the distance between the anchor sample and the positive sample (or the first distance). d(a,n) represents the distance between the negative sample pair, that is, the distance between the anchor sample and the negative sample (or the second distance). Margin is a hyperparameter that controls the difference in distance between positive and negative samples.

[0146] In some embodiments, the triplet loss function aims to minimize a first distance between the anchor sample and the positive sample and to maximize a second distance between the anchor sample and the negative sample to train the preference network model.

[0147] The preference network model outputs the singer preference features based on the singer's comprehensive features. The singer preference features are the singer's comprehensive features after dimensionality reduction. The goal of the above triplet loss function can also be understood as training the distance between singer preference features with similar relationships to be as small as possible, indicating that they are close in the feature vector space; training the distance between singer preference features that do not have similar relationships to be as large as possible, indicating that they are far away in the feature vector space.

[0148] In some embodiments, the difference value is calculated based on a distance metric to quantify the relationship between the singer preference features. For example, the difference value is calculated based on Euclidean distance or cosine similarity. This application is not limited to this.

[0149] In some embodiments, the distance between the singer preference features is characterized according to the difference value, the singer relative priority scores of different singers are obtained based on the difference value, and the similarity between the singer preference features of different singers is obtained based on the singer relative priority score. Optionally, the difference value between the singer preference features of different singers is calculated, and the difference value includes the first difference value between the singer preference features of the positive sample and the anchor sample in the positive sample pair according to the preference network model, and / or the second difference value between the singer preference features of the negative sample and the anchor sample in the negative sample pair according to the preference network model. Among them, the first difference value can represent the first distance between the anchor sample and the positive sample, and the second difference value can represent the second distance between the anchor sample and the negative sample.

[0150] In some embodiments, singer relative priority scores of different singers are generated based on the difference values, and similarities (or similarity relationships) between singer preference features of different singers are obtained based on the singer relative priority scores. Optionally, the singer relative priority scores are positively correlated with the singer preference features. The higher the singer relative priority scores, the higher the similarities between the singer preference features; the lower the singer relative priority scores, the lower the similarities between the singer preference features.

[0151] In some embodiments, the singer preference features can be sorted according to the singer relative priority score. For example, singer 1 corresponds to singer preference feature 1, and other singer preference features are sorted according to the singer relative priority score, from high to low, singer preference feature 2, singer preference feature 3, singer preference feature 4, etc. On the feature space vector, it can be considered that singer preference feature 2 is closest to singer preference feature 1, and singer preference feature 3 is second closest to singer preference feature 1, and so on. No further elaboration is given.

[0152] Optimizing the Preference Network Model

[0153] In some embodiments, the preference network model is optimized based on a similarity loss value, which refers to the loss value between the similarity between the singer preference features of different singers (predicted similarity) and the similarity between the attribute information of different singers (real similarity).

[0154] Optionally, obtain the similarity between singer preference features of different singers, and obtain the similarity between attribute information of different singers; determine the similarity loss value based on the similarity between singer preference features of different singers and the similarity between attribute information of different singers, and optimize the trained preference network model based on the similarity loss value between different singers.

[0155] In some embodiments, the similarity between attribute information of different singers is obtained based on the interactive behavior data (ie, data corresponding to the behavior information) of users (audience objects) to songs of different singers. The similarity between attribute information of different singers of multiple singers is introduced below.

[0156] In some embodiments, the similarity between the attribute information of different singers is obtained based on a graph neural network model. Optionally, a topological structure of multiple singers is obtained, and the node features of each node in the topological structure respectively represent the attribute information of the corresponding singer, and the attribute information includes the operational behavior information of the singer's audience object on the singer. Optionally, the topological structure of multiple singers and the node features of the corresponding nodes of each singer in the topological structure are input into a trained graph neural network model, and the graph neural network model outputs the similarity between the attribute information of different singers; wherein the node features are obtained based on the operational behavior information of the audience object on the singer.

[0157] Among them, each node in the topological structure represents a singer, and the input of the graph neural network model includes node features corresponding to multiple singers. The node features are used to characterize the features corresponding to the attribute information corresponding to the singer on each node in the topological structure. For example, the node features can be features constructed based on the operational behavior information of the singer's audience object (such as data corresponding to the playback behavior, data corresponding to the like behavior, data corresponding to the collection behavior, etc.). The above is only an illustrative example, and this application does not limit the node features of the graph neural network model input.

[0158] In some embodiments, the graph neural network model is described as GraphSAGE. The nodes in GraphSAGE represent different singers, the edges in GraphSAGE represent the similarities between different singers, the weights of the edges represent the degree of similarity between singers, and the similarity relationship output by GraphSAGE for different singers is the degree of similarity between different singers. Optionally, the similarity between singers is obtained based on the user's operation behavior information, for example, the similarity relationship between singers is obtained based on the playback behavior of a large number of users.

[0159] Among them, the graph neural network model is a network model trained based on the operational behavior information of the audience object. Exemplarily, when training the graph neural network model (GraphSAGE), a graph structure is first constructed, and the operational behavior information of the audience object is used to construct the graph structure. Each node in the graph structure can be a singer, and the edge represents the similarity between singers. For example, if the audience object collects or plays the songs of two singers in succession, then an edge can be established between the two singers. The similarity between singers can also be determined based on the operational behavior information of multiple audience objects. During the training process, for each target node, GraphSAGE randomly samples its neighbor nodes (i.e., similar other singers). GraphSAGE updates the representation of the target node by aggregating the features of the neighbor nodes (i.e., similar other singers). The purpose of the aggregation operation is to integrate the information of the neighbor nodes into the features of the target singer, combine the features of the sampled neighbor nodes with the features of the target node, and update the representation of the target node through a nonlinear activation function (e.g., ReLU). Then, the model parameters of GraphSAGE are optimized based on the loss function. For example, for a music recommendation system, a cross entropy loss function can be used to measure the prediction accuracy of GraphSAGE, and the model parameters are updated using a back propagation algorithm. The node representation in GraphSAGE is optimized by minimizing the loss function.

[0160] It should be noted that the above is only an exemplary description of the graph neural network model and does not constitute a limitation on the training of the graph neural network model. The embodiments of the present application only use the similarity of the graph neural network model output for different singers (or understand it as the similarity relationship between the attribute information of different singers in the graph neural network model).

[0161] In some embodiments, a similarity loss value is determined based on the similarity between singer preference features of different singers and the similarity between attribute information of different singers, and a trained preference network model is optimized based on the similarity loss values ​​between different singers. The optimized preference network model can learn the similarity relationship in the graph neural network model and output singer preference features with similarity relationships. For example, the singer preference features output by the preference network model have the following characteristics: singer preference features with high similarity relationships (i.e., high similarity) are close in distance in the feature vector space, and singer preference features with low similarity relationships (i.e., low similarity) are far apart in the feature vector space.

[0162] Recommend target singers to the recommendation object based on the preference feature set

[0163] In some embodiments, the trained preference network model outputs singer preference features with similar relationships, and a preference feature set can be constructed based on multiple singer preference features. In the subsequent recommendation process, the music recommendation system can recommend a target singer to the recommendation object based on the preference singer of the recommendation object and the preference feature set.

[0164] Figure 5 is a flowchart of a method for recommending a preferred network model provided by an exemplary embodiment of the present application. The method can be executed by a computer device, which can be Figure 1 The method comprises the following steps:

[0165] Step 310: Obtain the preferred singer of the recommendation object and obtain the preferred feature set;

[0166] Among them, the preference feature set is a set constructed by the singer preference features corresponding to each singer output by the trained preference network model.

[0167] In some embodiments, the trained preference network model is used to output singer preference features based on the singer comprehensive features. The singer comprehensive features are used to characterize the comprehensive information features of the singer. Optionally, the preference feature set is a set constructed by the singer preference features corresponding to each singer output by the trained preference network model, and the preference feature set includes multiple singer preference features.

[0168] In the preference feature set, the singer preference features corresponding to singers with similar relationships or high similarity relationships are close in distance on the feature vector space; the singer preference features corresponding to singers with no similarity relationships or low similarity relationships are far in distance on the feature vector space.

[0169] Step 320: Query the first singer preference feature corresponding to the preferred singer in the preference feature set;

[0170] In some embodiments, the music recommendation system obtains the preferred singer of the recommended object, and optionally obtains the preferred singer of the recommended object based on the behavior information of the recommended object, for example, the preferred singer of the recommended object is obtained according to the play or collection of the recommended object. Exemplarily, if the recommended object A plays Cantonese singer A 50 times, the music recommendation system determines that the recommended object A likes Cantonese singer A and sets Cantonese singer A as the preferred singer of the recommended object A.

[0171] Optionally, the preference feature set includes multiple singer preference features. After obtaining the preferred singer of the recommendation object, the music recommendation system searches the preference feature set for a first singer preference feature corresponding to the preferred singer.

[0172] Step 330: Determine a second singer preference feature in the preference feature set whose similarity to the first singer preference feature meets a preset condition, and recommend the singer corresponding to the second singer preference feature as the target singer to the recommendation object.

[0173] The similarity degree reaches the preset condition specifically when the feature vector space distance between the second singer preference feature and the first singer preference feature is close. For example, the second singer preference feature is a feature that has a similar relationship with the first singer preference feature, or the second singer preference feature is a feature that has a similar relationship with the first singer preference feature that reaches a third similarity threshold. Optionally, the second singer preference feature includes one or more, which is not limited in this application.

[0174] Among them, the singers corresponding to the second singer preference feature include popular singers and unpopular singers, that is, the popular singers and unpopular singers are recommended to the recommendation objects as target singers.

[0175] Exemplary, with reference to Figure 6, the recommended object is "Xiao Bai Tian Kai Xin", the music recommendation system obtains the preferred singer of the recommended object "Xiao Bai Tian Kai Xin", for example, the preferred singer is determined according to the collection volume of the recommended object "Xiao Bai Tian Kai Xin", the music recommendation system searches for the first singer preference feature corresponding to the preferred singer in the preference feature set, and determines the second singer preference feature in the preference feature set that is close to the first singer preference feature, for example, including 8 second singer preference features that have a similar relationship with the first singer preference feature, the music recommendation system recommends the singers corresponding to the 8 second singer preference features as target singers to the recommended object "Xiao Bai Tian Kai Xin", the target singers include popular singers and unpopular singers, for example, on the display interface 600 of the recommended object, 601 in the playlist is displayed as an unpopular singer, that is, "Singer 3" and "Singer 5" are unpopular singers.

[0176] In summary, the method provided in the embodiment of the present application is that the music recommendation system obtains the preferred singer of the recommendation object, determines the first singer preference feature corresponding to the preferred singer in the preference feature set according to the preferred singer of the recommendation object, and determines the second singer preference feature with a similar distance according to the first singer preference feature, and recommends the target singer corresponding to the second singer preference feature to the recommendation object. On the one hand, the music recommendation system is based on the preferred singer, and can quickly locate the singer who best matches the preferred singer in the huge preference feature set, thereby reducing the time cost of users searching for music content and improving the recommendation efficiency. On the other hand, the recommended singers include popular singers and unpopular singers, and this method improves the similarity recall of unpopular singers.

[0177] Figure 7 The structural block diagram of a training device for a preference network model provided by an embodiment of the present application is shown. The training device for the preference network model has the function of implementing the above-mentioned training method example of the preference network model, and the function can be implemented by hardware, or by hardware executing corresponding software. The training device can be the server described above, or it can be set in the server. Figure 7 As shown, the training device may include: a processing module 710 , a construction module 720 , an output module 730 and a training module 740 .

[0178] The processing module 710 is used to obtain the singer comprehensive features of each singer based on the singer audio features of the representative songs of each singer among the multiple singers and the singer meta information features of each singer;

[0179] A construction module 720 is used to construct a plurality of positive sample pairs and a plurality of negative sample pairs based on the similarity between the singer comprehensive features of different singers; wherein the positive sample pair includes the singer comprehensive features of two singers whose similarity meets a preset similarity condition, and the negative sample pair includes the singer comprehensive features of two singers whose difference meets a preset difference condition;

[0180] Output module 730, used for inputting positive sample pairs and negative sample pairs into the preference network model to be trained, so that the preference network model outputs the singer preference features of each singer; the spatial dimension of the singer comprehensive feature is higher than the spatial dimension of the singer preference feature;

[0181] The training module 740 is used to train the preference network model using the similarity between the singer preference features of different singers to obtain the trained preference network model.

[0182] Figure 8 The block diagram of the structure of the device for recommending a preference network model provided by one embodiment of the present application is shown. The device for recommending a preference network model has the function of implementing the example of the recommendation method of the preference network model described above, and the function can be implemented by hardware or by hardware executing corresponding software. The device for recommending can be the server described above, or it can be set in the server. Figure 8 As shown, the recommendation device may include: an acquisition module 810 , a query module 820 and a recommendation module 830 .

[0183] An acquisition module 810 is used to acquire the preferred singers of the recommendation object and to acquire a preferred feature set, wherein the preferred feature set is a set constructed by the singer preference features corresponding to each singer output by the trained preference network model, and the trained preference network is obtained by the training method of the preference network model;

[0184] A query module 820, configured to query the preference feature set for a first singer preference feature corresponding to the preferred singer;

[0185] The recommendation module 830 is used to determine a second singer preference feature in the preference feature set whose similarity with the first singer preference feature meets a preset condition, and recommend the singer corresponding to the second singer preference feature as a target singer to the recommendation object.

[0186] It should be noted that the specific limitations in the embodiments of the training device for one or more preference network models and the recommendation device for preference network models provided above can be found in the limitations of the training method for preference network models and the recommendation method for preference network models mentioned above, and will not be repeated here. Each module of the above-mentioned device can be implemented in whole or in part by software, hardware and a combination thereof. Each module can be embedded in the processor of a computer device in the form of hardware or be independent of it, or it can be stored in the memory of a computer device in the form of software, so that the processor can call and execute the operations corresponding to each module.

[0187] The serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0188] Fig. 9The structural block diagram of a computer device 1200 shown in an exemplary embodiment of the present application is shown. The computer device can be used to implement the training method of the preference network model or the recommendation method of the preference network model provided in the above embodiments. The computer device 1200 includes a central processing unit (CPU) 1201, a system memory 1204 including a random access memory (RAM) 1202 and a read-only memory (ROM) 1203, and a system bus 1205 connecting the system memory 1204 and the central processing unit 1201. The computer device 1200 also includes a basic input / output system (I / O system) 1206 that helps transmit information between various devices in the computer device, and a large-capacity storage device 1207 for storing an operating system 1213, an application program 1214 and other program modules 1215.

[0189] The basic input / output system 1206 includes a display 1208 for displaying information and an input device 1209 such as a mouse and a keyboard for user inputting information. The display 1208 and the input device 1209 are connected to the central processing unit 1201 through an input / output controller 1210 connected to the system bus 1205. The basic input / output system 1206 may also include an input / output controller 1210 for receiving and processing inputs from a plurality of other devices such as a keyboard, a mouse, or an electronic stylus. Similarly, the input / output controller 1210 also provides output to a display screen, a printer, or other types of output devices.

[0190] The mass storage device 1207 is connected to the central processing unit 1201 via a mass storage controller (not shown) connected to the system bus 1205. The mass storage device 1207 and its associated computer-readable storage medium provide non-volatile storage for the terminal device 1200. That is, the mass storage device 1207 may include a computer-readable storage medium (not shown) such as a hard disk or a compact disc read-only memory (CD-ROM) drive.

[0191] Without loss of generality, the computer-readable storage medium may include computer storage media and communication media. Computer storage media include volatile and non-volatile, removable and non-removable media implemented by any method or technology for storing information such as computer-readable storage instructions, data structures, program modules or other data. Computer storage media include RAM, ROM, Erasable Programmable Read Only Memory (EPROM), Electronically Erasable Programmable Read-Only Memory (EEPROM), flash memory or other solid-state storage technology, CD-ROM, Digital Versatile Disc (DVD) or other optical storage, cassette, tape, disk storage or other magnetic storage devices. Of course, those skilled in the art will know that the computer storage medium is not limited to the above. The above-mentioned system memory 1204 and mass storage device 1207 can be collectively referred to as memory.

[0192] The memory stores one or more programs, and the one or more programs are configured to be executed by one or more central processing units 1201. The one or more programs include instructions for implementing the above-mentioned method embodiments. The central processing unit 1201 executes the one or more programs to implement the training method of the preference network model provided by the above-mentioned various method embodiments, and / or the recommendation method of the preference network model.

[0193] According to various embodiments of the present application, the computer device 1200 can also be connected to a remote terminal device on a network through a network such as the Internet. That is, the computer device 1200 can be connected to the network 1212 through the network interface unit 1211 connected to the system bus 1205, or the network interface unit 1211 can be used to connect to other types of networks or remote terminal device systems (not shown).

[0194] The memory also includes one or more programs, which are stored in the memory and include steps for performing the method provided in the embodiment of the present application and executed by the terminal device.

[0195] An embodiment of the present application also provides a computer-readable storage medium, which stores at least one computer program, and the at least one computer program is loaded and executed by a processor to implement the training method of the preference network model provided in the above-mentioned method embodiments, and / or the recommendation method of the preference network model.

[0196] An embodiment of the present application also provides a computer program product, which includes a computer program, and the computer program is stored in a computer-readable storage medium; the computer program is read and executed from the computer-readable storage medium by a processor of a computer device, so that the computer device executes to implement the training method of the preference network model provided in the above-mentioned method embodiments, and / or the recommendation method of the preference network model.

[0197] It will be understood that in the specific implementation of the present application, the data involved, historical data, and data related to user data processing such as portraits related to user identity or characteristics, when the above embodiments of the present application are applied to specific products or technologies, need to obtain user permission or consent, and the collection, use and processing of relevant data need to comply with relevant laws, regulations and standards of relevant countries and regions.

[0198] It should be noted that, unless otherwise clearly defined herein, all terms used in the claims are interpreted according to their ordinary meanings in the technical field. Unless otherwise clearly stated, all references to "an element, device, component, equipment, step, etc." will be openly interpreted as referring to at least one instance of an element, device, component, equipment, step, etc. Unless explicitly stated, the steps of any method disclosed herein do not have to be performed in the exact order disclosed.

Claims

1. A training method for a preference network model, characterized in that: The method comprises: Based on the singer audio features of the representative songs of each singer among the multiple singers and the singer meta-information features of each singer, a singer comprehensive feature of each singer is obtained; Based on the similarity between the singer comprehensive features of different singers, a plurality of positive sample pairs and a plurality of negative sample pairs are constructed; wherein the positive sample pair includes the singer comprehensive features of two singers whose similarity meets a preset similarity condition, and the negative sample pair includes the singer comprehensive features of two singers whose difference meets a preset difference condition; Inputting the positive sample pair and the negative sample pair into the preference network model to be trained, so that the preference network model outputs the singer preference feature of each singer; the spatial dimension of the singer comprehensive feature is higher than the spatial dimension of the singer preference feature; The preference network model is trained using the similarities between the singer preference features of the different singers to obtain a trained preference network model.

2. The method according to claim 1, characterized in that The preference network model is trained based on a triplet loss function, and the preference network model is trained using the similarity between the singer preference features of different singers to obtain a trained preference network model, including: Determine the triplet loss function based on the positive sample pair and the negative sample pair; The preference network model is trained with the goal of minimizing a first distance between the comprehensive features of the two singers in the positive sample pair and maximizing a second distance between the comprehensive features of the two singers in the negative sample pair to obtain a trained preference network model; The first distance and the second distance are used to characterize the similarity between the singer preference features of the different singers.

3. The method according to claim 1, characterized in that The method further comprises: Obtaining the similarity between the singer preference features of the different singers, and obtaining the similarity between the attribute information of the different singers; A similarity loss value is determined based on the similarity between the singer preference features of the different singers and the similarity between the attribute information of the different singers, and the trained preference network model is optimized based on the similarity loss value.

4. The method according to claim 3, characterized in that The method further comprises: Acquire a topological structure of the plurality of singers, wherein the node feature of each node in the topological structure respectively represents the attribute information of each singer, and the attribute information includes the operation behavior information of the audience object on each singer; The topological structure of the multiple singers and the node features of the nodes corresponding to each singer in the topological structure are input into a trained graph neural network model, and the graph neural network model outputs the similarity between the attribute information of the different singers.

5. The method according to claim 1, characterized in that The singer comprehensive features of each singer are obtained based on the singer audio features of the representative songs of each singer among the multiple singers and the singer meta-information features of each singer, including: Inputting the singer audio features of each singer into a pre-trained audio alignment model, and outputting the singer audio features after alignment processing; Inputting the singer meta information features of each singer into a pre-trained meta information alignment model, and outputting the singer meta information features after alignment processing; wherein the singer audio features after alignment processing and the singer meta information features after alignment processing are features in the same spatial dimension; The singer audio features after the alignment processing and the singer meta-information features after the alignment processing are spliced ​​to obtain the singer comprehensive features of each singer.

6. The method according to any one of claims 1 to 5, characterized in that: The singer audio feature includes at least one of the following: a pitch feature of a representative song, a rhythm feature of a representative song, a speed feature of a representative song, and a melody feature of a representative song; The singer meta-information features include at least one of the following: the singer's gender feature, the singer's corresponding genre feature, the singer's representative song's language feature, and the singer's representative song's release year feature.

7. A recommendation method for a preference network model, characterized in that: The method comprises: Obtaining a preferred singer of a recommendation object and a preferred feature set, wherein the preferred feature set includes a set constructed by singer preference features of each singer output by a trained preference network model, wherein the preference network model is trained by the training method according to any one of claims 1 to 6; Querying the preference feature set for a first singer preference feature corresponding to the preferred singer; Determine a second singer preference feature in the preference feature set whose similarity to the first singer preference feature meets a preset condition, and recommend the singer corresponding to the second singer preference feature as a target singer to the recommendation object.

8. A computer device, characterized in that: The computer device includes: a processor and a memory, wherein at least one computer program is stored in the memory, and at least one computer program is loaded and executed by the processor to implement the training method of the preference network model as described in any one of claims 1 to 6, and the recommendation method of the preference network model as described in claim 7.

9. A computer storage medium, characterized in that The computer-readable storage medium stores at least one computer program, and the at least one computer program is loaded and executed by the processor to implement the training method of the preference network model as described in any one of claims 1 to 6, and the recommendation method of the preference network model as described in claim 7.

10. A computer program product, characterized in that The computer program product includes a computer program, which is stored in a computer-readable storage medium; the computer program is read and executed from the computer-readable storage medium by a processor of a computer device, so that the computer device executes the training method of the preference network model as described in any one of claims 1 to 6, and the recommendation method of the preference network model as described in claim 7.

Images