Training method of song recommendation model, computer equipment and readable storage medium

Abstract

The invention relates to a song recommendation model training method, computer equipment and a readable storage medium. The method comprises the following steps: acquiring an interactive song set of a historical putting object, and determining a plurality of first sample combinations and at least one type of second sample combinations according to the interactive song set; determining a first loss function value according to song representation information of each first sample combination output by the song recommendation model; determining a second loss function value according to song representation information of each second sample combination output by the song recommendation model; determining a third loss function value according to song representation information corresponding to song fragments in each first sample combination output by the song recommendation model; and training a song recommendation model according to the first loss function value, the second loss function value and the third loss function value. By adopting the method, limited song interaction historical information can be fully utilized for modeling, the performance of a song recommendation model is effectively improved, and the improvement of a new song putting effect is facilitated.

Description

Technical Field

[0001] The present application relates to the field of Internet technology, and in particular to a training method for a song recommendation model, a computer device, a computer-readable storage medium, and a computer program product. Background Art

[0002] At present, the traditional song recommendation system uses the user's long-term accumulated song listening history to construct a training data set for training. This song recommendation system is suitable for the recommendation scenario of ordinary songs, especially when a large number of users have already listened to the ordinary song.

[0003] For new song launch scenarios, it is impossible to infer the appropriate user group based on historical user behavior before promotion. It can only be analyzed and pushed based on the audio information of the song itself to solve the "cold start" problem of new songs. Since there is not enough user listening history to learn from, it is difficult to collect sufficient training samples. The traditional song recommendation method cannot fully utilize the existing information, and the limited user song interaction information is not fully mined, resulting in poor launch effects for new songs. Summary of the invention

[0004] Based on this, it is necessary to provide a training method, computer equipment, computer-readable storage medium and computer program product for a song recommendation model that can improve the effect of new song release in response to the above technical problems.

[0005] In a first aspect, the present application provides a method for training a song recommendation model. The method comprises:

[0006] Obtain an interactive song set of a historical delivery object, and determine a plurality of first sample combinations and at least one type of second sample combination according to the interactive song set; wherein the first sample combination includes two songs, the second sample combination includes at least three songs, and the number of songs included in the second sample combinations of the same type is the same; at least one song in different first sample combinations is different, and different second sample combinations in the same type of second sample combinations include at least three different songs;

[0007] Determine a first loss function value according to the song characterization information of each of the first sample combinations output by the song recommendation model; the first loss function value is used to characterize the characteristic distance between the songs in the first sample combination;

[0008] Determine a second loss function value according to the song characterization information of each second sample combination output by the song recommendation model; the second loss function value is used to characterize the characteristic distance between each song in the second sample combination and the preference center;

[0009] Determine a third loss function value according to the song characterization information corresponding to the song segments in each of the first sample combinations output by the song recommendation model; the third loss function value is used to characterize the characteristic distance between the song segments of the songs in the first sample combination;

[0010] The song recommendation model is trained according to the first loss function value, the second loss function value, and the third loss function value.

[0011] In a second aspect, the present application provides a song recommendation method. The method comprises:

[0012] Get the target song in the new song delivery task;

[0013] Determine a target song preference object that matches the target song from a plurality of song preference objects through a song recommendation model; wherein the song recommendation model is trained by any of the song recommendation model training methods;

[0014] The target song is sent to the target song preference object.

[0015] In a third aspect, the present application further provides a computer device. The computer device includes a memory and a processor, the memory stores a computer program, and the processor implements the steps of the song recommendation model training method as described in the first aspect and / or the steps of the song recommendation method as described in the second aspect when executing the computer program.

[0016] In a fourth aspect, the present application further provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the song recommendation model training method described in the first aspect and / or the steps of the song recommendation method described in the second aspect are implemented.

[0017] In a fifth aspect, the present application further provides a computer program product. The computer program product includes a computer program, which, when executed by a processor, implements the steps of the song recommendation model training method as described in the first aspect, and / or the steps of the song recommendation method as described in the second aspect.

[0018] The training method, computer device, computer-readable storage medium and computer program product of the song recommendation model described above obtain an interactive song collection of historical delivery objects, determine multiple first sample combinations and at least one type of second sample combination based on the interactive song collection, the first sample combination includes two songs, the second sample combination includes at least three songs, and the number of songs included in the same type of second sample combination is the same, at least one song in different first sample combinations is different, and different second sample combinations in the same type of second sample combination include at least three different songs, and then determine a first loss function value based on the song characterization information of each first sample combination output by the song recommendation model, the first loss function value is used to characterize the feature distance between songs in the first sample combination, and determine a second loss function value based on the song characterization information of each second sample combination output by the song recommendation model. The second loss function value is used to characterize the characteristic distance between each song in the second sample combination and the preference center. The third loss function value is determined according to the song characterization information corresponding to the song fragments in each first sample combination output by the song recommendation model. The third loss function value is used to characterize the characteristic distance between the song fragments of the songs in the first sample combination. Then, according to the first loss function value, the second loss function value, and the third loss function value, the song recommendation model is trained, thereby realizing the training optimization of the song recommendation model. The model can be trained based on the calculation method of characteristic distance loss according to the combination of songs in pairs, the combination of multiple songs, and the local slices of the combination of songs in pairs. It can make full use of the limited historical information of song interaction for modeling, and conduct a more detailed analysis of the situation of insufficient training data in the new song release scenario, thereby effectively improving the performance of the song recommendation model and helping to improve the effect of new song release. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] FIG1 is a schematic flow chart of a method for training a song recommendation model in one embodiment;

[0020] FIG2 is a schematic diagram of a song recommendation model training process in one embodiment;

[0021] FIG3 is a schematic diagram of a flow chart of a song recommendation method in one embodiment;

[0022] FIG4 is a flow chart of a method for training a song recommendation model in another embodiment;

[0023] FIG5 is a structural block diagram of a training device for a song recommendation model in one embodiment;

[0024] FIG6 is a structural block diagram of a song recommendation device in one embodiment;

[0025] FIG7 is a diagram showing the internal structure of a computer device in one embodiment;

[0026] FIG. 8 is a diagram showing the internal structure of a computer device in another embodiment. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0028] In an exemplary embodiment, as shown in FIG1 , a method for training a song recommendation model is provided. This embodiment uses the method applied to a terminal as an example. It is understandable that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. The server can be implemented as an independent server or a server cluster composed of multiple servers; the terminal can be, but is not limited to, various personal computers, laptops, smart phones, tablet computers, portable wearable devices, etc.

[0029] In this embodiment, the method includes the following steps S101 to S104. Among them:

[0030] In step S101, an interactive song set of a historical delivery object is obtained, and a plurality of first sample combinations and at least one type of second sample combination are determined according to the interactive song set.

[0031] The first sample combination may include two songs, the second sample combination may include at least three songs, and the same type of second sample combination may include the same number of songs.

[0032] As an example, different first sample combinations may include at least one different song, that is, both songs may be different, or only one song may be different, and different second sample combinations in the same type of second sample combination may include at least three different songs.

[0033] In actual applications, as shown in Figure 2, the song list list of users who have released new songs (i.e., historical release objects) can be obtained for screening, such as the user's favorite song list, the user's complete play song list, etc., to obtain an interactive song collection, and then the songs in the interactive song collection can be combined, and the obtained combination can be filtered, and the filtered combination can be used as training data.

[0034] Specifically, the first sample combination can be obtained by combining any two songs in pairs, and the second sample combination can include a random combination of three songs, a random combination of four songs, or a random combination of five songs, or other random song combinations, which are not specifically limited in this embodiment. In this way, the training samples can be fully utilized when the training data is insufficient.

[0035] In step S102, a first loss function value is determined according to the song characterization information of each first sample combination output by the song recommendation model, wherein the first loss function value can be used to characterize the characteristic distance between songs in the first sample combination.

[0036] As an example, the song representation information of the first sample combination may include a song audio representation and a song text representation of each song in the first sample combination.

[0037] In a specific implementation, the song recommendation model may include an audio representation extraction model and a text representation extraction model. As shown in FIG2 , each song in each first sample combination may be respectively input into the audio representation extraction model MERT (Music Understanding Model with Large-Scale Self-supervised Training) and the text representation extraction model BERT (Bidirectional Encoder Representations from Transformers) to obtain the song audio representation and song text representation of each song in each first sample combination as the song representation information of the first sample combination, and then the ternary loss may be calculated according to the song representation information of each first sample combination to obtain the first loss function value.

[0038] In step S103, a second loss function value is determined according to the song characterization information of each second sample combination output by the song recommendation model, wherein the second loss function value can be used to characterize the characteristic distance between each song in the second sample combination and the preference center.

[0039] As an example, the song representation information of the second sample combination may include a song audio representation and a song text representation of each song in the second sample combination.

[0040] Exemplarily, as shown in FIG2 , each song in each second sample combination can be respectively input into the audio representation extraction model MERT and the text representation extraction model BERT to obtain the song audio representation and song text representation of each song in each second sample combination as the song representation information of the second sample combination, and then the central loss can be calculated according to the song representation information of each second sample combination to obtain the second loss function value.

[0041] In step S104, a third loss function value is determined according to the song characterization information corresponding to the song segments in each first sample combination output by the song recommendation model, wherein the third loss function value can be used to characterize the characteristic distance between the song segments of the songs in the first sample combination.

[0042] As an example, song segments of two songs in each first sample combination can be obtained by random slicing to obtain a combined segment, and the song representation information corresponding to the song segments in the first sample combination may include song audio representation and song text representation of the song segments of the two songs in the combined segment.

[0043] In practical applications, as shown in Figure 2, the song clips in each first sample combination obtained by random slicing can be respectively input into the audio representation extraction model MERT and the text representation extraction model BERT to obtain the song audio representation and song text representation of each song clip as the song representation information corresponding to the song clip in the first sample combination, and then the distance loss of the slice can be calculated according to the song representation information corresponding to the song clip in each first sample combination to obtain the third loss function value.

[0044] In step S105, a song recommendation model is trained according to the first loss function value, the second loss function value, and the third loss function value.

[0045] In a specific implementation, the target loss function value can be obtained according to the first loss function value, the second loss function value, and the third loss function value, and then the song recommendation model can be trained by the target loss function value until the training end condition is met to obtain a trained song recommendation model; as shown in Figure 2, the model training can be based on the resnet deep residual network, or neural networks of other structures can be used, which are not specifically limited in this embodiment.

[0046] For example, the first loss function value, the second loss function value, and the third loss function value can be added to obtain the target loss function value, and then the model can be trained until the target loss function value converges, and the model can be saved as a recommendation model associated with song preferences, that is, a trained song recommendation model. In this way, the user interaction song information of the existing delivery behavior can be fully utilized, the user preferences can be deeply mined, and a song recommendation system with better performance can be built.

[0047] In the training method of the above-mentioned song recommendation model, by obtaining the interactive song set of the historical delivery object, multiple first sample combinations and at least one type of second sample combination are determined according to the interactive song set, and then the first loss function value is determined according to the song representation information of each first sample combination output by the song recommendation model, the second loss function value is determined according to the song representation information of each second sample combination output by the song recommendation model, and the third loss function value is determined according to the song representation information corresponding to the song fragments in each first sample combination output by the song recommendation model. Then, according to the first loss function value, the second loss function value, and the third loss function value, the song recommendation model is trained, and the training optimization of the song recommendation model is realized. The model can be trained according to the calculation method of feature distance loss based on pairwise combinations of songs and multiple song combinations, as well as local slices of pairwise combinations of songs, which can make full use of the limited historical information of song interactions for modeling, and conduct a more detailed analysis of the situation of insufficient training data in the new song delivery scenario, effectively improving the performance of the song recommendation model and helping to improve the effect of new song delivery.

[0048] In an exemplary embodiment, in step S101, obtaining an interactive song set of a historical delivery object, and determining a plurality of first sample combinations and at least one type of second sample combination according to the interactive song set may include the following steps:

[0049] Obtain an interactive song collection of historical delivery objects, the interactive song collection being a song collection that meets a preset length screening condition; randomly combine songs in the interactive song collection to obtain a plurality of first song combinations and at least one type of second song combination corresponding to the interactive song collection, wherein the first song combination includes two songs, the second song combination includes at least three songs, and the number of songs included in the second song combination of the same type is the same; according to the occurrence frequency of each first song combination, filter from the plurality of first song combinations a first song combination whose occurrence frequency meets a first preset frequency threshold as a first sample combination; and according to the occurrence frequency of each second song combination, filter from the same type of second song combination a second song combination whose occurrence frequency meets a second preset frequency threshold as a second sample combination.

[0050] In actual applications, as shown in Figure 2, it is possible to filter from the user's favorite song list associated with the delivery behavior (i.e., the song collection of the historical delivery object), and then filter out the collection song list (i.e., the interactive song collection) whose song list length is greater than 5 (i.e., the preset length filter condition of the song). Among them, by selecting the song list associated with the delivery behavior, it can be ensured that the trained model is more suitable for the delivery scenario; the data with less than 5 songs in the song list can be judged as untrustworthy data and filtered, that is, the collection song list that is not in the confidence interval is filtered out. Of course, the song length filter condition can also be other values, not limited to the number shown in Figure 2.

[0051] In one example, the songs in the interactive song set can be combined and the frequency of occurrence of each combination can be counted. Optionally, any two songs can be combined in pairs, three songs can be randomly combined, four songs can be randomly combined, and five songs can be randomly combined in the same interactive song set, and the frequency (number of times) of occurrence of each combination can be counted. Then, the combination can be filtered according to the statistical frequency to retain the high-confidence combination, and then the filtered combination can be used as training data. Among them, the combination of two songs can be called the first song combination, and the combination of three songs, the combination of four songs, and the combination of five songs can be called the second song combination. The first sample combination can be obtained by screening from the first song combination, and the second sample combination can be obtained by screening from the second song combination. It should be noted that the second sample combination can include fewer categories or more categories, such as the second sample combination can include only one category, that is, a combination of three songs, or the second sample combination can include four categories, that is, a combination of three songs, a combination of four songs, a combination of five songs, and a combination of six songs.

[0052] It can be understood that if the interactive song sets obtained by filtering through the song filtering conditions include multiple ones, each interactive song set can be combined in the above manner to obtain the first sample combination and the second sample combination.

[0053] In the scenario of new song release, due to the small amount of user historical data, analysis based on the song dimension cannot fully utilize the information. By adopting a more detailed analysis method, a more detailed analysis can be performed based on the song associations of user interactions (songs in the same song list have preference associations, such as user collection preferences, user playback preferences, etc.), thereby fully learning user preference associations using limited data.

[0054] In an optional embodiment, different thresholds can be used to filter song combinations of different lengths (including different numbers of songs) to obtain a first sample combination and a second sample combination; as shown in Figure 2, the frequency threshold can be set to 1 for each song combination (i.e., the first song combination) to obtain the first sample combination, and the frequency thresholds can be set to 2, 3, and 4 for each three-song combination, four-song combination, and five-song combination (i.e., the second song combination) to obtain the second sample combination.

[0055] For example, taking a song collection of a historical delivery object as an example (including 9 songs A, B, C, D, E, F, G, H, and I), for the songs in the song collection, any two songs are combined as the first song combination, and any three songs are randomly combined, four songs are randomly combined, and five songs are randomly combined as the second song combination. It can be set to stop the combination processing after generating a certain number of random combinations or stop the combination processing based on specific conditions (such as the total duration of songs in all combinations meets the conditions). Then, the frequency of occurrence of each combination is counted and filtered using the frequency threshold to obtain the first sample combination and the second sample combination. As shown in Figure 2, assuming that among the multiple first song combinations obtained by the combination, the song combinations AB and CD both reach a frequency of 1 occurrence, then the song combinations AB and CD can be used as the first sample combination; assuming that among the multiple second song combinations obtained by the combination, the song combinations ABE and CDF both reach a frequency of 2 occurrences (such as 1 occurrence of ABE, 1 occurrence of BAE; 1 occurrence of CDF, 1 occurrence of FCD), then the song combinations ABE and CDF can be used as the second sample combination. Similarly, the song combination ABGH reaching a frequency of 3 occurrences can also be used as the second sample combination, and the song combination ABGHI reaching a frequency of 4 occurrences can also be used as the second sample combination. It can be understood that other song combinations that do not meet the threshold frequency are filtered out and cannot be called the first sample combination and the second sample combination. As shown in Figure 2, after filtering, the song combinations ABE, CDF, ABGH, and ABGHI (i.e., the second sample combination) are obtained.

[0056] In this embodiment, in response to the situation where training data is insufficient in the new song release scenario, based on a more detailed analysis, it is possible to fully learn the user preference correlation using limited data.

[0057] In an exemplary embodiment, the following steps may also be included:

[0058] Through the audio representation extraction model in the song recommendation model, the song audio representation of each song in each first sample combination and the song audio representation of each song in each second sample combination are obtained; through the text representation extraction model in the song recommendation model, the song text representation of each song in each first sample combination and the song text representation of each song in each second sample combination are obtained; the song audio representation and song text representation of each song in each first sample combination are used as the song representation information of the first sample combination, and the song audio representation and song text representation of each song in each second sample combination are used as the song representation information of the second sample combination.

[0059] As an example, the audio representation extraction model may be a MERT model (Music Understanding Model with Large-Scale Self-supervised Training), or other models may be used to extract the audio representation of a song; the text representation extraction model may be a BERT model (Bidirectional Encoder Representations from Transformers), or other models may be used to extract the text representation of a song.

[0060] In a specific implementation, as shown in FIG2 , the entire song and the entire lyrics of each song in each first sample combination and each second sample combination can be obtained and respectively input into the MERT model and the BERT model to extract the corresponding song audio representation and song text representation.

[0061] In this embodiment, the audio representation extraction model in the song recommendation model is used to obtain the song audio representation of each song in each first sample combination and the song audio representation of each song in each second sample combination. The text representation extraction model in the song recommendation model is used to obtain the song text representation of each song in each first sample combination and the song text representation of each song in each second sample combination. Then, the song audio representation and song text representation of each song in each first sample combination are used as the song representation information of the first sample combination, and the song audio representation and song text representation of each song in each second sample combination are used as the song representation information of the second sample combination. This can make full use of the user interactive song information of the existing delivery behavior, deeply mine the song representation, and provide data support for further model training.

[0062] In an exemplary embodiment, in step S102, determining the first loss function value according to the song representation information of each first sample combination output by the song recommendation model may include the following steps:

[0063] For each first sample combination among the multiple first sample combinations, the first audio loss of the multiple first sample combinations is determined based on the distance between the song audio representations of the two songs in each first sample combination and the distance between the song audio representations of any song in each first sample combination and any song in any other first sample combination outside each first sample combination; the first text loss of the multiple first sample combinations is determined based on the distance between the song text representations of the two songs in each first sample combination and the distance between the song text representations of any song in each first sample combination and any song in any other first sample combination outside each first sample combination; and the first loss function value is obtained according to the first audio loss and the first text loss.

[0064] In one example, for the entire song audio in any two song combinations (i.e., any first sample combination), the audio-related ternary loss (which may be referred to as the first audio loss) may be calculated in the following manner:

[0065]

[0066] Among them, Ad(a, p) represents the distance between the audio representations of two whole songs in a pairwise song combination, Ad(a, n) is the distance between a song randomly selected from other pairwise song combinations other than the pairwise song combination and any song in the pairwise song combination, and margin is used to control the size of the loss function; Loss1 is the first audio loss of any first sample combination, and the first audio losses of all first sample combinations are obtained by summing and averaging the first audio losses of each first sample combination.

[0067] For the entire song lyrics in the pairwise song combination, the text-related ternary loss (also called the first text loss) can be calculated as follows:

[0068]

[0069] Among them, Loss2 is the first text loss of any first sample combination, and the first text losses of each first sample combination are summed and averaged to obtain the first text losses of all first sample combinations. It should be noted that the explanation of each parameter can be found in the relevant description of the above-mentioned first audio loss formula. The difference is that the distance in the calculation formula of the first text loss is the distance between texts.

[0070] In this embodiment, for each first sample combination in the multiple first sample combinations, the first audio loss of the multiple first sample combinations is determined based on the distance between the song audio representations of the two songs in each first sample combination and the distance between the song audio representations of any song in each first sample combination and any song in any other first sample combination outside each first sample combination, and then the first text loss of the multiple first sample combinations is determined based on the distance between the song text representations of the two songs in each first sample combination and the distance between the song text representations of any song in each first sample combination and any song in any other first sample combination outside each first sample combination, and then the first loss function value is obtained according to the first audio loss and the first text loss, so as to bring two songs with preference association closer in the feature space.

[0071] In an exemplary embodiment, in step S103, determining the second loss function value according to the song representation information of each second sample combination output by the song recommendation model may include the following steps:

[0072] For each second sample combination in at least one category of second sample combinations, the second audio loss of at least one category of second sample combinations is determined based on the distance between the song audio representation of each song in each second sample combination and the song audio representation of the preference center in each second sample combination; the song audio representation of the preference center is the representation mean of the song audio representations of the songs in each second sample combination; the second text loss of at least one category of second sample combinations is determined based on the distance between the song text representation of each song in each second sample combination and the song text representation of the preference center in each second sample combination; the song text representation of the preference center is the representation mean of the song text representations of the songs in each second sample combination; the second loss function value is obtained according to the second audio loss and the second text loss.

[0073] In one example, for the entire song audio in a combination other than the pairwise song combination (ie, each second sample combination), the audio-related center loss (which may be referred to as the second audio loss) may be calculated in the following manner:

[0074]

[0075] Among them, Aemb mean Aemb represents the mean of the audio representation of all songs in each second sample combination (such as three songs, four songs, and five songs). n is the number of songs in the combination. When three songs are in a group, n is 3, when four songs are in a group, n is 4, and when five songs are in a group, n is 5. iRepresents the audio representation of each song in each second sample combination; Loss3 is the second audio loss of any second sample combination, and the second audio losses of each second sample combination are summed and averaged to obtain the second audio losses of all second sample combinations.

[0076] For the entire song lyrics in combinations other than the pairwise song combination, the text-related central loss (also called the second text loss) can be calculated as follows:

[0077]

[0078] Among them, Loss4 is the second text loss of any second sample combination, and the second text losses of each second sample combination are summed and averaged to obtain the second text losses of all second sample combinations. It should be noted that the explanation of each parameter can be found in the relevant description of the above second audio loss formula. The difference is that the distance in the calculation formula of the second text loss is the distance between texts.

[0079] In another example, by calculating the distance between the songs in the combination and the center through the central loss, the distance between the songs in the combination can be shortened. Compared with the ternary loss calculation method of the combination of two songs, the central loss calculation method can be more stable, reliable and more preferred.

[0080] In this embodiment, for each second sample combination in at least one category of second sample combinations, the second audio loss of at least one category of second sample combinations is determined based on the distance between the song audio representation of each song in each second sample combination and the song audio representation of the preference center in each second sample combination, and then the second text loss of at least one category of second sample combinations is determined based on the distance between the song text representation of each song in each second sample combination and the song text representation of the preference center in each second sample combination, and then the second loss function value is obtained according to the second audio loss and the second text loss. For the second sample combination, a preference center can be established through local sampling, so that the preference feature space is not too discrete, which helps to obtain an effective song recommendation model.

[0081] In an exemplary embodiment, in step S104, determining the third loss function value according to the song representation information corresponding to the song segments in each first sample combination output by the song recommendation model may include the following steps:

[0082] For each first sample combination in multiple first sample combinations, obtain at least one song fragment combination corresponding to each first sample combination; wherein the song fragment combination includes a song fragment of each of the two songs in the first sample combination; based on the distance between the song audio representations of the two song fragments in all the song fragment combinations in at least one song fragment combination, determine the third audio loss of the multiple first sample combinations; based on the distance between the song text representations of the two song fragments in all the song fragment combinations in at least one song fragment combination, determine the third text loss of the multiple first sample combinations; according to the third audio loss and the third text loss, obtain a third loss function value.

[0083] In practical applications, in addition to the information of the entire song, in order to make full use of limited training samples, local information of the song can be obtained by random slicing; for example, two songs in a pairwise combination can be randomly sliced ​​M times (one slice obtains a set of pairwise combination fragments), and M sets of pairwise combination song fragments are obtained. Then, the audio distance loss of the song fragments can be calculated in the following way:

[0084]

[0085] Among them, seg j1 and seg j2 are audio representations of two song clips respectively. Loss5 is the third audio loss of any song clip combination obtained based on multiple first sample combinations, and the third audio loss of each song clip combination is summed and averaged to obtain the third audio loss of all song clip combinations.

[0086] For the lyrics segment corresponding to the song segment, the text distance loss of the lyrics segment can be calculated as follows:

[0087]

[0088] Among them, Loss6 is the third text loss of any song segment combination obtained based on multiple first sample combinations, and the third text losses of each song segment combination are summed and averaged to obtain the third text loss of all song segment combinations. It should be noted that the explanation of each parameter can be found in the relevant description of the third audio loss formula above. The difference is that the distance in the calculation formula of the third text loss is the distance between texts.

[0089] Exemplarily, the above-mentioned Loss1-Loss6 losses can be added together to obtain a total loss function value to train the model.

[0090] In one example, the technical solution of this embodiment, by expanding the preferred songs in pairs to multiple preferred related song combinations, expanding the ternary loss to the center loss, and introducing the distance loss by means of local slicing, can combine the song audio and lyrics text, and make full use of the limited user song interaction history information for modeling, which can more carefully and effectively deeply explore user preferences and build a song recommendation model with better performance.

[0091] In another example, the technical solution of this embodiment, for the first sample combination, can bring two songs with preference associations closer together in the feature space, and perform local sampling on the associated songs so that the lyrics and songs are closer in the feature space; for the second sample combination, multiple songs with preference associations can be local sampled to establish preference centers, which can make the preferences more representative and the preference feature space not too discrete, which is helpful to quickly build models using a small amount of training data when little training data is available, and obtain an effective song recommendation system.

[0092] In this embodiment, for the first sample combination, local sampling of related songs can be achieved, so that the lyrics and songs are closer in the feature space.

[0093] In an exemplary embodiment, as shown in FIG3 , a song recommendation method is provided, which is described by taking the method applied to a server as an example, and includes the following steps S301 to S303. Among them:

[0094] In step S301, the target song in the new song delivery task is obtained.

[0095] In step S302, a target song preference object matching the target song is determined from multiple song preference objects through a song recommendation model; wherein the song recommendation model is trained by any song recommendation model training method.

[0096] In step S303, the target song is sent to the target song preference object.

[0097] In practical applications, the song recommendation system is an important way for users to listen to songs and release new songs. By recommending suitable songs to suitable users, it can help users expand their playlists and discover treasure songs. Song recommendation is an important means to increase the breadth of music. The technical solution of this embodiment analyzes the situation of insufficient training data in the new song release scenario of song recommendation, and proposes a song recommendation method that is effectively applied to the release scenario. By adopting a song recommendation model with better performance that fully utilizes the user interactive song information of existing release behaviors, it can recommend target songs to matching target song preference objects for new song release tasks, which can effectively improve the effect of new song release.

[0098] In the above song recommendation method, by obtaining the target song in the new song delivery task, and then using the song recommendation model to determine the target song preference object that matches the target song from multiple song preference objects, and then sending the target song to the target song preference object, the new song delivery effect can be effectively improved based on the optimized song recommendation model.

[0099] In one embodiment, as shown in FIG4 , a flow chart of another method for training a song recommendation model is provided. In this embodiment, the method includes the following steps:

[0100] In step S401, the interactive song set of the historical delivery object is obtained, and multiple first sample combinations and at least one type of second sample combination are determined according to the interactive song set. In step S402, the audio representation of each song in each first sample combination and the audio representation of each song in each second sample combination are obtained through the audio representation extraction model in the song recommendation model. In step S403, the song text representation of each song in each first sample combination and the song text representation of each song in each second sample combination are obtained through the text representation extraction model in the song recommendation model. In step S404, the song audio representation and song text representation of each song in each first sample combination are used as the song representation information of the first sample combination, and the song audio representation and song text representation of each song in each second sample combination are used as the song representation information of the second sample combination. In step S405, the first loss function value is determined according to the song representation information of each first sample combination output by the song recommendation model. In step S406, the second loss function value is determined according to the song representation information of each second sample combination output by the song recommendation model. In step S407, the third loss function value is determined according to the song representation information corresponding to the song segments in each first sample combination output by the song recommendation model. In step S408, the song recommendation model is trained according to the first loss function value, the second loss function value, and the third loss function value. It should be noted that the specific limitations of the above steps can be found in the specific limitations of the training method of a song recommendation model above, which will not be repeated here.

[0101] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0102] Based on the same inventive concept, the embodiment of the present application also provides a training device for a song recommendation model for implementing the training method for the song recommendation model involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in the embodiments of the training device for one or more song recommendation models provided below can refer to the limitations of the training method for the song recommendation model above, and will not be repeated here.

[0103] In one embodiment, as shown in FIG5 , a training device for a song recommendation model is provided, comprising:

[0104] The training data determination module 501 is used to obtain an interactive song set of a historical delivery object, and determine a plurality of first sample combinations and at least one type of second sample combination according to the interactive song set; wherein the first sample combination includes two songs, the second sample combination includes at least three songs, and the number of songs included in the second sample combination of the same type is the same; at least one song in different first sample combinations is different, and different second sample combinations in the same type of second sample combination include at least three different songs;

[0105] A first loss function value determination module 502 is used to determine a first loss function value according to the song characterization information of each of the first sample combinations output by the song recommendation model; the first loss function value is used to characterize the characteristic distance between the songs in the first sample combination;

[0106] A second loss function value determination module 503 is used to determine a second loss function value according to the song characterization information of each second sample combination output by the song recommendation model; the second loss function value is used to characterize the characteristic distance between each song in the second sample combination and the preference center;

[0107] A third loss function value determination module 504 is used to determine a third loss function value according to the song characterization information corresponding to the song segments in each of the first sample combinations output by the song recommendation model; the third loss function value is used to characterize the characteristic distance between the song segments of the songs in the first sample combination;

[0108] The model training module 505 is used to train the song recommendation model according to the first loss function value, the second loss function value, and the third loss function value.

[0109] In one embodiment, the training data determination module 501 includes:

[0110] An interactive song set determination submodule, used to obtain an interactive song set of the historical delivery object, wherein the interactive song set is a song set that meets a preset length screening condition;

[0111] A song combination obtaining submodule is used to randomly combine the songs in the interactive song set to obtain a plurality of first song combinations and at least one type of second song combination corresponding to the interactive song set, wherein the first song combination includes two songs, the second song combination includes at least three songs, and the number of songs included in the same type of the second song combination is the same;

[0112] A combination filtering submodule is used to filter, according to the occurrence frequency of each of the first song combinations, from the multiple first song combinations to obtain a first song combination whose occurrence frequency meets a first preset frequency threshold as a first sample combination; and, according to the occurrence frequency of each of the second song combinations, to filter, from the same type of the second song combinations, to obtain a second song combination whose occurrence frequency meets a second preset frequency threshold as a second sample combination.

[0113] In one embodiment, the apparatus further comprises:

[0114] An audio representation extraction model processing module, used to obtain a song audio representation of each song in each of the first sample combinations and a song audio representation of each song in each of the second sample combinations through the audio representation extraction model in the song recommendation model;

[0115] A text representation extraction model processing module, used to obtain a song text representation of each song in each of the first sample combinations and a song text representation of each song in each of the second sample combinations through a text representation extraction model in the song recommendation model;

[0116] The sample combination information obtaining module is used to use the song audio representation and song text representation of each song in each of the first sample combinations as the song representation information of the first sample combination, and to use the song audio representation and song text representation of each song in each of the second sample combinations as the song representation information of the second sample combination.

[0117] In one embodiment, the first loss function value determination module 502 includes:

[0118] A first audio loss determination submodule is used to determine, for each first sample combination of the plurality of first sample combinations, a first audio loss of the plurality of first sample combinations based on a distance between song audio representations of two songs in the plurality of first sample combinations, and a distance between song audio representations of any song in the plurality of first sample combinations and any song in any other plurality of first sample combinations other than the plurality of first sample combinations;

[0119] A first text loss determination submodule, configured to determine the first text loss of the plurality of first sample combinations based on the distance between the song text representations of the two songs in each first sample combination and the distance between the song text representations of any song in each first sample combination and any song in any other first sample combination other than the first sample combination;

[0120] The first loss function value obtaining submodule is used to obtain the first loss function value according to the first audio loss and the first text loss.

[0121] In one embodiment, the second loss function value determination module 503 includes:

[0122] A second audio loss determination submodule is used to determine, for each second sample combination in the at least one category of second sample combinations, a second audio loss of the at least one category of second sample combinations based on the distance between the song audio representation of each song in the at least one category of second sample combinations and the song audio representation of the preference center in the at least one category of second sample combinations; the song audio representation of the preference center is the representation mean of the song audio representations of the songs in the at least one category of second sample combinations;

[0123] A second text loss determination submodule is configured to determine the second text loss of the at least one type of second sample combination based on the distance between the song text representation of each song in each second sample combination and the song text representation of the preference center in each second sample combination; the song text representation of the preference center is the representation mean of the song text representations of the songs in each second sample combination;

[0124] The second loss function value obtaining submodule is used to obtain the second loss function value according to the second audio loss and the second text loss.

[0125] In one embodiment, the third loss function value determination module 504 includes:

[0126] A slicing submodule, configured to obtain, for each first sample combination in the plurality of first sample combinations, at least one song segment combination corresponding to the first sample combination; wherein the song segment combination includes a song segment of each of the two songs in the first sample combination;

[0127] A third audio loss determination submodule, configured to determine a third audio loss of the plurality of first sample combinations based on a distance between song audio representations of two song segments in all of the song segment combinations in the at least one song segment combination;

[0128] A third text loss determination submodule, which determines a third text loss of the plurality of first sample combinations based on a distance between song text representations of two song segments in all the song segment combinations in the at least one song segment combination;

[0129] The third loss function value obtaining submodule is used to obtain the third loss function value according to the third audio loss and the third text loss.

[0130] Each module in the training device of the song recommendation model can be implemented in whole or in part by software, hardware and a combination thereof. Each module can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute operations corresponding to each module.

[0131] Based on the same inventive concept, the embodiment of the present application also provides a song recommendation device for implementing the song recommendation method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more song recommendation device embodiments provided below can refer to the limitations of the song recommendation method above, and will not be repeated here.

[0132] In one embodiment, as shown in FIG6 , a song recommendation device is provided, comprising:

[0133] The target song acquisition module 601 is used to acquire the target song in the new song delivery task;

[0134] A target song preference object determination module 602 is used to determine a target song preference object that matches the target song from a plurality of song preference objects through a song recommendation model; wherein the song recommendation model is trained by any of the song recommendation model training methods;

[0135] The song delivery module 603 is used to deliver the target song to the target song preference object.

[0136] Each module in the above song recommendation device can be implemented in whole or in part by software, hardware or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute the operations corresponding to each module.

[0137] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be shown in FIG7. The computer device includes a processor, a memory, an input / output interface, a communication interface, a display unit, and an input device. The processor, the memory, and the input / output interface are connected via a system bus, and the communication interface, the display unit, and the input device are connected to the system bus via the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be implemented through WIFI, a mobile cellular network, NFC (near field communication) or other technologies. When the computer program is executed by the processor, a training method for a song recommendation model is implemented.

[0138] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be shown in FIG8. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. The processor, the memory and the input / output interface are connected via a system bus, and the communication interface is connected to the system bus via the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store song recommendation data. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a song recommendation method is implemented.

[0139] Those skilled in the art will understand that the structures shown in FIGS. 7-8 are merely block diagrams of partial structures related to the scheme of the present application, and do not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device may include more or fewer components than shown in the figures, or combine certain components, or have a different arrangement of components.

[0140] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:

[0141] Obtain an interactive song set of a historical delivery object, and determine a plurality of first sample combinations and at least one type of second sample combination according to the interactive song set; wherein the first sample combination includes two songs, the second sample combination includes at least three songs, and the number of songs included in the second sample combinations of the same type is the same; at least one song in different first sample combinations is different, and different second sample combinations in the same type of second sample combinations include at least three different songs;

[0142] Determine a first loss function value according to the song characterization information of each of the first sample combinations output by the song recommendation model; the first loss function value is used to characterize the characteristic distance between the songs in the first sample combination;

[0143] Determine a second loss function value according to the song characterization information of each second sample combination output by the song recommendation model; the second loss function value is used to characterize the characteristic distance between each song in the second sample combination and the preference center;

[0144] Determine a third loss function value according to the song characterization information corresponding to the song segments in each of the first sample combinations output by the song recommendation model; the third loss function value is used to characterize the characteristic distance between the song segments of the songs in the first sample combination;

[0145] The song recommendation model is trained according to the first loss function value, the second loss function value, and the third loss function value.

[0146] In one embodiment, when the processor executes the computer program, the steps in the other embodiments described above are also implemented.

[0147] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:

[0148] Get the target song in the new song delivery task;

[0149] Determine a target song preference object that matches the target song from a plurality of song preference objects through a song recommendation model; wherein the song recommendation model is trained by any of the song recommendation model training methods;

[0150] The target song is sent to the target song preference object.

[0151] In one embodiment, when the processor executes the computer program, the steps in the other embodiments described above are also implemented.

[0152] In one embodiment, a computer readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the following steps are implemented:

[0153] Obtain an interactive song set of a historical delivery object, and determine a plurality of first sample combinations and at least one type of second sample combination according to the interactive song set; wherein the first sample combination includes two songs, the second sample combination includes at least three songs, and the number of songs included in the second sample combinations of the same type is the same; at least one song in different first sample combinations is different, and different second sample combinations in the same type of second sample combinations include at least three different songs;

[0154] Determine a first loss function value according to the song characterization information of each of the first sample combinations output by the song recommendation model; the first loss function value is used to characterize the characteristic distance between the songs in the first sample combination;

[0155] Determine a second loss function value according to the song characterization information of each second sample combination output by the song recommendation model; the second loss function value is used to characterize the characteristic distance between each song in the second sample combination and the preference center;

[0156] Determine a third loss function value according to the song characterization information corresponding to the song segments in each of the first sample combinations output by the song recommendation model; the third loss function value is used to characterize the characteristic distance between the song segments of the songs in the first sample combination;

[0157] The song recommendation model is trained according to the first loss function value, the second loss function value, and the third loss function value.

[0158] In one embodiment, when the computer program is executed by a processor, the steps in the other embodiments described above are also implemented.

[0159] In one embodiment, a computer readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the following steps are implemented:

[0160] Get the target song in the new song delivery task;

[0161] Determine a target song preference object that matches the target song from a plurality of song preference objects through a song recommendation model; wherein the song recommendation model is trained by any of the song recommendation model training methods;

[0162] The target song is sent to the target song preference object.

[0163] In one embodiment, when the computer program is executed by a processor, the steps in the other embodiments described above are also implemented.

[0164] In one embodiment, a computer program product is provided, comprising a computer program, which, when executed by a processor, implements the following steps:

[0165] Obtain an interactive song set of a historical delivery object, and determine a plurality of first sample combinations and at least one type of second sample combination according to the interactive song set; wherein the first sample combination includes two songs, the second sample combination includes at least three songs, and the number of songs included in the second sample combinations of the same type is the same; at least one song in different first sample combinations is different, and different second sample combinations in the same type of second sample combinations include at least three different songs;

[0166] Determine a first loss function value according to the song characterization information of each of the first sample combinations output by the song recommendation model; the first loss function value is used to characterize the characteristic distance between the songs in the first sample combination;

[0167] Determine a second loss function value according to the song characterization information of each second sample combination output by the song recommendation model; the second loss function value is used to characterize the characteristic distance between each song in the second sample combination and the preference center;

[0168] Determine a third loss function value according to the song characterization information corresponding to the song segments in each of the first sample combinations output by the song recommendation model; the third loss function value is used to characterize the characteristic distance between the song segments of the songs in the first sample combination;

[0169] The song recommendation model is trained according to the first loss function value, the second loss function value, and the third loss function value.

[0170] In one embodiment, when the computer program is executed by a processor, the steps in the other embodiments described above are also implemented.

[0171] In one embodiment, a computer program product is provided, comprising a computer program, which, when executed by a processor, implements the following steps:

[0172] Get the target song in the new song delivery task;

[0173] Determine a target song preference object that matches the target song from a plurality of song preference objects through a song recommendation model; wherein the song recommendation model is trained by any of the song recommendation model training methods;

[0174] The target song is sent to the target song preference object.

[0175] In one embodiment, when the computer program is executed by a processor, the steps in the other embodiments described above are also implemented.

[0176] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application 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 relevant laws, regulations and standards of relevant countries and regions.

[0177] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited to this.

[0178] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0179] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A method for training a song recommendation model, characterized in that: The method comprises: Obtain an interactive song set of a historical delivery object, and determine a plurality of first sample combinations and at least one type of second sample combination according to the interactive song set; wherein the first sample combination includes two songs, the second sample combination includes at least three songs, and the number of songs included in the second sample combinations of the same type is the same; at least one song in different first sample combinations is different, and different second sample combinations in the same type of second sample combinations include at least three different songs; Determine a first loss function value according to the song characterization information of each of the first sample combinations output by the song recommendation model; the first loss function value is used to characterize the characteristic distance between the songs in the first sample combination; Determine a second loss function value according to the song characterization information of each second sample combination output by the song recommendation model; the second loss function value is used to characterize the characteristic distance between each song in the second sample combination and the preference center; Determine a third loss function value according to the song characterization information corresponding to the song segments in each of the first sample combinations output by the song recommendation model; the third loss function value is used to characterize the characteristic distance between the song segments of the songs in the first sample combination; The song recommendation model is trained according to the first loss function value, the second loss function value, and the third loss function value.

2. The method according to claim 1, characterized in that The step of acquiring an interactive song set of a historical delivery object and determining a plurality of first sample combinations and at least one type of second sample combination according to the interactive song set includes: Obtaining an interactive song set of the historical delivery object, wherein the interactive song set is a song set that meets a preset length screening condition; Randomly combining the songs in the interactive song set to obtain a plurality of first song combinations and at least one type of second song combination corresponding to the interactive song set, wherein the first song combination includes two songs, the second song combination includes at least three songs, and the number of songs included in the same type of the second song combination is the same; According to the occurrence frequency of each of the first song combinations, first song combinations whose occurrence frequency meets a first preset frequency threshold are filtered out from the multiple first song combinations as first sample combinations; and according to the occurrence frequency of each of the second song combinations, second song combinations whose occurrence frequency meets a second preset frequency threshold are filtered out from the same type of the second song combinations as second sample combinations.

3. The method according to claim 1, characterized in that The method further comprises: Obtaining, by means of the audio representation extraction model in the song recommendation model, a song audio representation of each song in each of the first sample combinations and a song audio representation of each song in each of the second sample combinations; Obtaining a song text representation of each song in each of the first sample combinations and a song text representation of each song in each of the second sample combinations through a text representation extraction model in the song recommendation model; The song audio representation and song text representation of each song in each of the first sample combinations are used as the song representation information of the first sample combination, and the song audio representation and song text representation of each song in each of the second sample combinations are used as the song representation information of the second sample combination.

4. The method according to claim 3, characterized in that The determining of a first loss function value according to the song representation information of each of the first sample combinations output by the song recommendation model includes: For each first sample combination of the plurality of first sample combinations, determining a first audio loss of the plurality of first sample combinations based on a distance between song audio representations of two songs in the each first sample combination and a distance between song audio representations of any song in the each first sample combination and any song in any other first sample combination other than the each first sample combination; Determine the first text loss of the plurality of first sample combinations based on the distance between the song text representations of the two songs in each first sample combination and the distance between the song text representations of any song in each first sample combination and any song in any other first sample combination other than each first sample combination; The first loss function value is obtained according to the first audio loss and the first text loss.

5. The method according to claim 3, characterized in that: The determining of the second loss function value according to the song representation information of each second sample combination output by the song recommendation model includes: For each second sample combination in the at least one category of second sample combinations, the second audio loss of the at least one category of second sample combinations is determined based on the distance between the song audio representation of each song in the each second sample combination and the song audio representation of the preference center in the each second sample combination; the song audio representation of the preference center is the representation mean of the song audio representations of the songs in the each second sample combination; Determine the second text loss of the at least one type of second sample combination based on the distance between the song text representation of each song in each second sample combination and the song text representation of the preference center in each second sample combination; the song text representation of the preference center is the representation mean of the song text representations of the songs in each second sample combination; The second loss function value is obtained according to the second audio loss and the second text loss.

6. The method according to claim 3, characterized in that The determining of the third loss function value according to the song representation information corresponding to the song segments in each of the first sample combinations output by the song recommendation model includes: For each first sample combination in the plurality of first sample combinations, obtaining at least one song segment combination corresponding to each first sample combination; wherein the song segment combination includes a song segment of each of the two songs in the first sample combination; determining a third audio loss of the plurality of first sample combinations based on a distance between the song audio representations of two song segments in all the song segment combinations in the at least one song segment combination; determining a third text loss of the plurality of first sample combinations based on a distance between song text representations of two song segments in all the song segment combinations in the at least one song segment combination; The third loss function value is obtained according to the third audio loss and the third text loss.

7. A song recommendation method, characterized in that: The method comprises: Get the target song in the new song delivery task; Determine a target song preference object that matches the target song from a plurality of song preference objects through a song recommendation model; wherein the song recommendation model is trained by the song recommendation model training method according to any one of claims 1 to 6; The target song is sent to the target song preference object.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, it implements the training method of the song recommendation model described in any one of claims 1 to 6, and / or the steps of the song recommendation method described in claim 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the training method of the song recommendation model described in any one of claims 1 to 6, and / or the steps of the song recommendation method described in claim 7.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, it implements the training method of the song recommendation model described in any one of claims 1 to 6, and / or the steps of the song recommendation method described in claim 7.

Images