Song delivery model training method, song delivery method, and related device

Abstract

The application relates to a song delivery model training method, a song delivery method and related equipment. The method comprises the following steps: obtaining a training sample, a song matching model used for predicting the matching degree between a sample user and a sample song in the training sample, and a song rough arrangement model used for roughly arranging the pairing combination of the sample user and the sample song in the training sample; sampling a target sample from first samples filtered from the training sample by the song matching model and second samples filtered from the training sample by the song rough arrangement model; taking the training sample as training data of the song matching model and the song rough arrangement model, taking the target sample and the training sample as training data of a song fine arrangement model, jointly training the song matching model, the song rough arrangement model and the song fine arrangement model, and obtaining the trained song fine arrangement model. The method can improve the matching degree between a user and a delivered song and improve the song delivery efficiency.

Description

Technical Field

[0001] This application relates to the field of Internet technology, and in particular to a training method for a song delivery model, a song delivery method, a computer device, and a computer-readable storage medium. Background Technology

[0002] In music platform promotion scenarios, traditional technologies typically employ modeling methods from other fields (such as advertising and recommendations). However, song promotion differs from advertising and recommendations due to the unique nature of its business scenarios, making it unsuitable for direct application. Alternatively, a coarse-grained approach may be used, primarily focusing on deciding on the distribution channels and volume, resulting in low matching between specific users and songs and impacting the efficiency of song promotion.

[0003] Therefore, there is a problem with the poor performance of song delivery in the relevant technologies. Summary of the Invention

[0004] Therefore, it is necessary to provide a training method, computer equipment, and computer-readable storage medium for a song delivery model that can improve the matching degree between users and delivered songs, in order to address the above-mentioned technical problems.

[0005] Firstly, this application provides a method for training a song distribution model. The method includes:

[0006] The system includes: acquiring training samples; a song matching model for predicting the matching degree between sample users and sample songs in the training samples; and a song ranking model for coarsely ranking the pairing combinations of sample users and sample songs in the training samples.

[0007] Target samples are sampled from the first sample selected by the song matching model based on the training samples and the second sample selected by the song coarse ranking model based on the training samples; the matching degree between the sample users and the sample songs in the target samples is less than the matching threshold, and the coarse ranking of the pairing combinations of sample users and sample songs in the target samples is greater than the ranking threshold.

[0008] The training samples are used as training data for the song matching model and the song coarse ranking model, and the target samples and the training samples are used as training data for the song fine ranking model; the song matching model, the song coarse ranking model and the song fine ranking model are jointly trained to obtain the trained song fine ranking model; the trained song fine ranking model is used as the song distribution model.

[0009] In one embodiment, before the step of sampling the target sample from the first sample selected by the song matching model based on the training samples and the second sample selected by the song ranking model based on the training samples, the method further includes:

[0010] Based on multiple sample users and multiple sample songs in the training samples, multiple constructed samples are obtained;

[0011] The matching prediction results of each constructed sample are obtained using the song matching model, and the first sample is selected from the plurality of constructed samples; the matching prediction results are used to characterize the degree of matching between the sample user and the sample song in the corresponding constructed sample;

[0012] The ranking prediction results of each constructed sample are obtained using the song coarse ranking model, and the second sample is selected from the multiple constructed samples; the ranking prediction results are used to characterize the coarse ranking of the constructed samples that are paired and combined by sample users and sample songs.

[0013] In one embodiment, obtaining the matching prediction results of each of the constructed samples using the song matching model, and selecting the first sample from the plurality of constructed samples, includes:

[0014] The sample user features in each of the constructed samples are input into the user-side model of the song matching model, and the sample song features in each of the constructed samples are input into the song-side model of the song matching model.

[0015] Based on the output results of the user-side model and the song-side model in the song matching model, the matching prediction results of each constructed sample are obtained.

[0016] The constructed sample corresponding to the matching prediction result that is less than the matching threshold is used as the first sample.

[0017] In one embodiment, obtaining the ranking prediction results of each constructed sample using the song coarse ranking model, and selecting the second sample from the plurality of constructed samples, includes:

[0018] The sample user features in each of the constructed samples are input into the user-side model of the song coarse ranking model, and the sample song features in each of the constructed samples are input into the song-side model of the song coarse ranking model.

[0019] Based on the target results output by the user-side model and the target results output by the song-side model in the coarse-grained song ranking model, the ranking prediction results of each constructed sample are obtained; the target result is the output vector obtained by taking the type of the positive sample as the learning target.

[0020] The constructed sample corresponding to the ranking prediction result that is greater than the ranking threshold is used as the second sample.

[0021] In one embodiment, sampling the target sample from the first sample selected by the song matching model based on the training samples and the second sample selected by the song coarse ranking model based on the training samples includes:

[0022] Multiple candidate samples are obtained based on the intersection of the first sample and the second sample;

[0023] The sampling quantity and preset sampling information are determined; the preset sampling information is used to indicate the sampling method determined based on the matching prediction results and ranking prediction results of each candidate sample.

[0024] The target sample is obtained by sampling the multiple candidate samples according to the sampling quantity and the preset sampling information.

[0025] In one embodiment, the method further includes:

[0026] Based on the target sample, the positive and negative samples in the training sample, a training sample set is generated for the song ranking model;

[0027] The user features and song features of each sample in the training sample set are concatenated to obtain the concatenated feature vector corresponding to each sample in the training sample set.

[0028] The concatenated feature vector corresponding to any sample in the training sample set is input into the song ranking model to obtain the model output result; the model output result is used to characterize the probability information that any sample belongs to the type of the positive sample, and the probability information is used to determine whether to perform the song distribution operation.

[0029] In one embodiment, the step of using the training samples as training data for the song matching model and the song coarse ranking model, and the target samples and the training samples as training data for the song fine ranking model, to jointly train the song matching model, the song coarse ranking model, and the song fine ranking model includes:

[0030] Obtain the first loss of the song matching model and the second loss of the song coarse ranking model when the training samples are used to train the song matching model and the song coarse ranking model;

[0031] Obtain the third loss of the song re-ranking model when training the song re-ranking model using the target sample and the training sample;

[0032] The first loss, the second loss, and the third loss are weighted and fused together, and the resulting fusion loss is used to update the parameters of the song matching model, the song coarse ranking model, and the song fine ranking model.

[0033] Secondly, this application provides a song delivery method, the method comprising:

[0034] When it is detected that the user to be targeted meets the song targeting trigger conditions, the target song is determined from multiple candidate songs by the song targeting model; wherein, the song targeting model is trained by the training method of the song targeting model described in any one of the first aspects;

[0035] The target song will be sent to the users to be targeted.

[0036] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the training method for the song delivery model as described in the first aspect, and / or the steps of the song delivery method as described in the second aspect.

[0037] Fourthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the steps of the training method for the song delivery model as described in the first aspect, and / or the steps of the song delivery method as described in the second aspect.

[0038] Fifthly, this application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, implements the steps of the training method for the song delivery model as described in the first aspect, and / or the steps of the song delivery method as described in the second aspect.

[0039] The aforementioned training method, song delivery method, computer device, and computer-readable storage medium for a song delivery model involve acquiring training samples, a song matching model for predicting the matching degree between sample users and sample songs in the training samples, and a song coarse ranking model for coarsely ranking the pairings of sample users and sample songs in the training samples. Then, target samples are sampled from a first sample selected by the song matching model based on the training samples and a second sample selected by the song coarse ranking model based on the training samples. In these target samples, the matching degree between sample users and sample songs is less than a matching threshold, and the coarse ranking of the pairings of sample users and sample songs in the target samples is greater than a ranking threshold. By using training samples as training data for the song matching model and the coarse-ranking model, and using target samples and training samples as training data for the fine-ranking model, the song matching model, the coarse-ranking model, and the fine-ranking model are jointly trained to obtain the trained fine-ranking model. This trained fine-ranking model is then used as the song delivery model, achieving training optimization for the song delivery model. Based on splitting the model into different stages for training and jointly training each stage, and selecting target samples to fuse with training samples to train the fine-ranking model, it can better balance the matching degree and ranking target between users and delivered songs, improving the matching degree between users and delivered songs, and also improving the efficiency of song delivery. Attached Figure Description

[0040] Figure 1 This is a flowchart illustrating a training method for a song delivery model in one embodiment.

[0041] Figure 2a This is a schematic diagram of a matching stage processing flow in one embodiment;

[0042] Figure 2b This is a schematic diagram of a coarse-sorting stage processing flow in one embodiment;

[0043] Figure 2c This is a schematic diagram of a sampling stage processing flow in one embodiment;

[0044] Figure 2d This is a schematic diagram of a fine-sorting stage processing flow in one embodiment;

[0045] Figure 3 This is a schematic diagram of the training process of a song rehearsal model in one embodiment;

[0046] Figure 4 This is a flowchart illustrating a song delivery method in one embodiment;

[0047] Figure 5 This is a flowchart illustrating the training method of the song delivery model in another embodiment;

[0048] Figure 6 This is a structural block diagram of a training device for a song delivery model in one embodiment;

[0049] Figure 7 This is a structural block diagram of a song delivery device in one embodiment;

[0050] Figure 8 This is an internal structural diagram of a computer device in one embodiment;

[0051] Figure 9 This is a diagram of the internal structure of a computer device in another embodiment. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0053] In one embodiment, such as Figure 1 As shown, a training method for a song delivery model is provided. This embodiment illustrates the application of this method to a server. It is understood that this method can also be applied to terminals, and can also be applied to systems including terminals and servers, and implemented through the interaction between the terminal and the server. The server can be a standalone server or a server cluster composed of multiple servers; the terminal can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices.

[0054] In this embodiment, the method includes the following steps S101 to S103. Wherein:

[0055] In step S101, training samples, a song matching model for predicting the matching degree between sample users and sample songs in the training samples, and a song coarse ranking model for coarsely ranking the pairing combinations of sample users and sample songs in the training samples are obtained.

[0056] As an example, the song matching model and the song coarse ranking model can be a dual-tower structure model, which can include a User tower (i.e., user-side model) and an Item tower (i.e., song-side model). The input of the User tower can be sample user features, and the input of the Item tower can be sample song features.

[0057] In practical applications, algorithm modeling can be performed for different stages during model training. For the matching stage, a song matching model to be trained can be constructed based on a dual-tower structure and trained using the contrastive learning Sampled Softmax method. For example, the matching score between sample users and sample songs in the training samples can be calculated, which can be used to predict the degree of matching between songs and user interests.

[0058] For example, for the coarse ranking stage, a dual-tower structure can be used to construct the coarse ranking model of the song to be trained. The learning objectives of the model can be set to include multiple types (such as Positive Sample, Negative Sample, Bad Sample), and the Softmax Loss method can be used for training to predict...<user,Item> The probability that a pairing of a sample user and a sample song is a positive example is used as the coarse ranking.

[0059] In step S102, the target sample is sampled from the first sample selected by the song matching model based on the training samples and the second sample selected by the song ranking model based on the training samples.

[0060] Among them, the matching degree between sample users and sample songs in the target sample is less than the matching threshold, and the coarse ranking of the pairing combinations of sample users and sample songs in the target sample is greater than the ranking threshold.

[0061] In the specific implementation, based on the song matching model, pairs with matching scores below a certain value (i.e., matching threshold) can be used as a candidate set for constructing Bad Samples, denoted as CandiBadSamplesByMatch (i.e., the first sample selected by the song matching model); based on the song coarse ranking model, pairs with coarse ranking scores above a certain value (i.e., ranking threshold) can be used as a candidate set for constructing Bad Samples, denoted as CandiBadSamplesByPreRank (i.e., the second sample selected by the song coarse ranking model).

[0062] In one example, for the sampling phase, the intersection of the first and second samples selected in the matching and coarse ranking phases can be used to obtain the Bad Sample candidate set, denoted as CandiBadSamplesCandi. Then, sampling can be performed according to the preset sampling strategy and sampling ratio to obtain Bad Samples (i.e., target samples).

[0063] In step S103, the training samples are used as training data for the song matching model and the song coarse ranking model, and the target samples and training samples are used as training data for the song fine ranking model. The song matching model, the song coarse ranking model and the song fine ranking model are jointly trained to obtain the trained song fine ranking model.

[0064] The trained song ranking model can be used as a song distribution model.

[0065] In practical applications, for the fine-ranking stage, a song fine-ranking model to be trained can be constructed based on a deep neural network. By optimizing the feature learning and feature interaction methods, and setting learning objectives including multiple types (such as Positive Sample, Negative Sample, Bad Sample) for training, the fine-ranking scores of users and items can be predicted. Then, the song matching model, the song coarse-ranking model, and the song fine-ranking model can be jointly trained to obtain the trained song fine-ranking model. The goal of training the song fine-ranking model is to improve the efficiency of the delivery task.

[0066] In one optional embodiment, the trained song ranking model can be used as a song distribution model, which can be applied to music platforms to promote and distribute songs. It is applicable to both professionally generated content (PGC) and user-generated content (UGC), and can be used during song promotion stages, such as when a new song is released or to help singers promote and distribute their songs, in order to help content creators gain more traffic exposure. It is not limited to specific channels, locations, or formats, such as official playlists, AI playlists, and editorial playlists.

[0067] Compared to traditional methods, song delivery differs from advertising and recommendations. It requires adding songs to the source song list, and while the delivered content shares the same source as recommendations, it is more novel or less popular. This makes the business scenario of song delivery unique. The technical solution in this embodiment employs an end-to-end intelligent song delivery matching method, systematically considering the matching problem between users and delivered songs. For example, it can be divided into matching, coarse ranking, sampling, and fine ranking stages to address the matching and ranking issue. By breaking it down into different stages for algorithm modeling, each stage can have different modeling goals and focuses. The matching, coarse ranking, and sampling stages serve the fine ranking stage, providing target samples for training. Furthermore, the stages can be integrated for joint training to balance the modeling goals of each stage.

[0068] The training method of the aforementioned song delivery model involves acquiring training samples, a song matching model for predicting the matching degree between sample users and sample songs in the training samples, and a song coarse ranking model for coarsely ranking the pairing combinations of sample users and sample songs in the training samples. Then, target samples are sampled from the first sample selected by the song matching model based on the training samples and the second sample selected by the song coarse ranking model based on the training samples. The training samples are then used as training data for the song matching model and the song coarse ranking model, and the target samples and training samples are used as training data for the song fine ranking model. The song matching model, the song coarse ranking model, and the song fine ranking model are jointly trained to obtain the trained song fine ranking model. This achieves training optimization for the song delivery model. By splitting the model into different stages for training and jointly training each stage, and selecting target samples to fuse with training samples to train the song fine ranking model, it can better balance the matching degree and ranking target between users and delivered songs, improve the matching degree between users and delivered songs, and also improve the efficiency of song delivery.

[0069] Specifically, among the probability scores of Positive Samples output by the model (i.e., the model output results), the songs with the highest probability scores are selected as songs that can be distributed. Bad Samples (i.e., target samples) are selected by combining the matching stage and the coarse ranking stage. For example, samples with low matching scores but high coarse ranking scores are not actually very suitable, even though they are ranked high. By learning to distinguish and suppress target samples in the fine ranking stage, a balance can be achieved between matching scores and ranking scores, that is, balancing the degree of matching between users and the songs to be distributed and the ranking target.

[0070] In one embodiment, before the step of sampling the target sample from the first sample selected based on the training samples by the song matching model and the second sample selected based on the training samples by the song coarse ranking model, the following step may be included:

[0071] Based on multiple sample users and multiple sample songs in the training samples, multiple constructed samples are obtained; a song matching model is used to obtain the matching prediction results of each constructed sample, and the first sample is selected from the multiple constructed samples; the matching prediction results are used to characterize the degree of matching between the sample users and sample songs in the corresponding constructed samples; a song coarse ranking model is used to obtain the ranking prediction results of each constructed sample, and the second sample is selected from the multiple constructed samples; the ranking prediction results are used to characterize the coarse ranking of the constructed samples composed of sample users and sample songs.

[0072] In one example, the sample users and sample songs in the training samples can be paired combinations, such as...<user,Item> It can combine multiple sample users and multiple sample songs, such as using a Cartesian product to construct new ones.<user,Item> This means that there are multiple constructed samples. Then, a song matching model and a song coarse ranking model can be used to process these multiple constructed samples respectively to filter out the first sample and the second sample.

[0073] In another example, the matching phase, coarse ranking phase, and sampling phase serve the fine ranking phase. These three phases can be used to provide bad samples for the fine ranking phase to train the model. The matching phase mainly learns the degree of matching between users and songs, while the coarse ranking phase mainly learns the coarse ranking of users and songs.

[0074] In this embodiment, multiple constructed samples are obtained based on multiple sample users and multiple sample songs in the training samples. A song matching model is used to obtain the matching prediction results of each constructed sample. A first sample is selected from the multiple constructed samples. A song coarse ranking model is used to obtain the ranking prediction results of each constructed sample. A second sample is selected from the multiple constructed samples, which provides data support for further sampling of target samples.

[0075] In one embodiment, using a song matching model to obtain the matching prediction results of each constructed sample, and selecting the first sample from multiple constructed samples, may include the following steps:

[0076] The user features in each constructed sample are input into the user-side model of the song matching model, and the song features in each constructed sample are input into the song-side model of the song matching model. Based on the output results of the user-side model and the song-side model of the song matching model, the matching prediction results of each constructed sample are obtained. The constructed sample corresponding to the matching prediction result that is less than the matching threshold is taken as the first sample.

[0077] In practical applications, the matching stage mainly involves calculating the matching score between the user and the song. A dual-tower model can be trained using contrastive learning, which allows the embedding distance of similar (same category) samples to be as close as possible, and the embedding distance of samples from different categories to be as far as possible.

[0078] In one example, such as Figure 2a As shown, the matching phase can take the following steps:

[0079] 1. New samples can be constructed based on InBatch batch data (i.e. training samples) to obtain Batch * Batch number * batch number (i.e., multiple constructed samples).

[0080] 2. Features can be input through a dual-tower model based on a dual-tower structure. For example, the User tower (i.e., the user-side model) takes user-side features (i.e., sample user features) as input, and the Item tower (i.e., the song-side model) takes song-side features (i.e., sample song features) as input.

[0081] 3. The inner product of the last layer output of the User tower and the Item tower can be taken as the Match Score (i.e. the matching prediction result).

[0082] 4. In the Batch*Batch samples, samples with a Match Score lower than a preset threshold can be used as CandiBadSamplesByMatch, that is, the constructed samples corresponding to the matching prediction results that are less than the matching threshold are used as the first samples.

[0083] 5. The loss can be calculated using Sampled Softmax, which means training the song matching model with training samples to obtain the first loss of the song matching model.

[0084] In an alternative embodiment, the matching stage can also be based on the idea of ​​user graph recall or other I2I-based (calculating the similarity between two items) approaches, and no specific restrictions are imposed in this embodiment.

[0085] In this embodiment, the user features of each constructed sample are input into the user-side model of the song matching model, and the song features of each constructed sample are input into the song-side model of the song matching model. Then, based on the output results of the user-side model and the song-side model of the song matching model, the matching prediction results of each constructed sample are obtained. The constructed samples corresponding to the matching prediction results that are less than the matching threshold are used as the first samples, providing a candidate dataset for subsequent sampling target samples.

[0086] In one embodiment, using a song coarse-ranking model to obtain the ranking prediction results of each constructed sample, and then selecting a second sample from multiple constructed samples, may include the following steps:

[0087] The user features from each constructed sample are input into the user-side model of the song coarse ranking model, and the song features from each constructed sample are input into the song-side model of the song coarse ranking model. Based on the target results output by the user-side model and the song-side model of the song coarse ranking model, the ranking prediction results of each constructed sample are obtained. The target result is the output vector obtained by taking the type of the positive sample as the learning target. The constructed sample corresponding to the ranking prediction result that is greater than the ranking threshold is used as the second sample.

[0088] In the specific implementation, the coarse ranking stage mainly calculates the coarse ranking scores of users and songs. It can also be based on the dual-tower structure model, and the learning objectives can be set as Positive Sample, Negative Sample, and Bad Sample.

[0089] In one example, such as Figure 2b As shown, the coarse sorting stage can be performed using the following steps:

[0090] 1. New samples can be constructed from InBatch batch data (i.e. training samples) to obtain Batch * Batch number * batch number (i.e., multiple constructed samples).

[0091] 2. Features can be input through a dual-tower model based on a dual-tower structure. For example, the User tower (i.e., the user-side model) takes user-side features (i.e., sample user features) as input, and the Item tower (i.e., the song-side model) takes song-side features (i.e., sample song features) as input. Each tower in the last layer has 3 Output Vectors, which correspond to the 3 learning objectives mentioned above.

[0092] 3. The Output Vector (i.e., target result) of the Positive Sample target (i.e., positive sample type) in the last layer of the User Tower and Item Tower can be taken, and the Inner Product can be calculated as the PreRank Score (i.e. ranking prediction result). The Inner Product of the Output Vectors of the other two targets can also be calculated to obtain the corresponding scores, which can be used to calculate the loss of the model.

[0093] 4. In the Batch*Batch samples, samples with PreRank Score higher than a preset threshold can be used as CandiBadSamplesByPreRank, that is, the constructed samples corresponding to the ranking prediction results that are greater than the ranking threshold are used as the second samples.

[0094] 5. Softmax Loss can be used to calculate the loss. For example, first calculate the Softmax of the above scores, and then calculate the loss. That is, when training the song coarse ranking model with training samples, the second loss of the song coarse ranking model is obtained.

[0095] In an optional embodiment, the coarse ranking stage may also use a threshold method that is not limited to a pre-set numerical threshold, but may also use the top k% method, or combine the score distribution, use discrete numerical thresholds, etc. No specific restrictions are imposed in this embodiment.

[0096] In this embodiment, the user features of each constructed sample are input into the user-side model of the song coarse ranking model, and the song features of each constructed sample are input into the song-side model of the song coarse ranking model. Then, based on the target results output by the user-side model and the target results output by the song-side model of the song coarse ranking model, the ranking prediction results of each constructed sample are obtained. The constructed samples corresponding to the ranking prediction results that are greater than the ranking threshold are used as the second samples, providing a candidate dataset for subsequent sampling target samples.

[0097] In one embodiment, in step S102, sampling the target sample from the first sample selected by the song matching model based on the training samples and the second sample selected by the song coarse ranking model based on the training samples may include the following steps:

[0098] Multiple candidate samples are obtained based on the intersection of the first and second samples; the sampling quantity and preset sampling information are determined; the preset sampling information is used to indicate the sampling method determined based on the matching prediction results and ranking prediction results of each candidate sample; the multiple candidate samples are sampled according to the sampling quantity and preset sampling information to obtain the target sample.

[0099] In practical applications, the sampling stage is mainly used to obtain Bad Samples (i.e., sample target samples) for use in the subsequent fine ranking stage. For example, samples with matching scores below a certain threshold but coarse ranking scores above a certain threshold (i.e., the matching degree between sample users and sample songs is less than the matching threshold, and the coarse ranking of the pairing combination of sample users and sample songs is greater than the ranking threshold) can be sampled to obtain Bad Samples (such as samples with low matching scores but high coarse ranking scores).

[0100] In one example, such as Figure 2c As shown, the sampling phase can be performed using the following steps:

[0101] 1. The intersection of CandiBadSamplesByMatch and CandiBadSamplesByPreRank can be used to obtain samples with low matching scores and high ranking scores, which are then used as CandiBadSamples. In other words, multiple candidate samples can be obtained based on the intersection of the first and second samples.

[0102] 2. You can set a preset sampling rate (e.g., 0.5) to get the number of samples within a batch, such as Number = Batch * Sampling Rate;

[0103] 3. Based on the Match Score and PreRank Score of each sample obtained by taking the intersection, the sampling probability (i.e., sampling distribution information) can be calculated, such as...

[0104]

[0105] 4. Then, based on the sampling probability and the number of samples, the Bad Sample, i.e. the target sample, can be obtained by sampling without replacement.

[0106] In an optional embodiment, other sampling methods may also be used in the sampling stage, such as sampling by sorting or sampling by matching; other strategies may also be set for the distribution of samples to make the final sampled Bad Samples more uniform and of higher quality, which are not specifically limited in this embodiment.

[0107] In this embodiment, multiple candidate samples are obtained by the intersection of the first sample and the second sample. Then, the sampling quantity and sampling distribution information are determined. Subsequently, the multiple candidate samples are sampled according to the sampling quantity and preset sampling information to obtain the target sample. This helps to balance the matching degree and ranking target between users and the songs being delivered, and also improves the training efficiency of the song delivery model.

[0108] In one embodiment, the following steps may also be included:

[0109] Based on the target sample, positive samples, and negative samples in the training sample, a training sample set for the song ranking model is generated. The user features and song features of each sample in the training sample set are concatenated to obtain the concatenated feature vector corresponding to each sample in the training sample set. The concatenated feature vector corresponding to any sample in the training sample set is input into the song ranking model to obtain the model output. The model output is used to characterize the probability information of any sample belonging to the type of the positive sample. The probability information is used to determine whether to perform the song placement operation.

[0110] In practice, the fine-ranking stage aims to further improve the matching degree between users and songs, thereby improving the efficiency of song delivery. On the other hand, similar to the coarse-ranking model, it can set learning targets as PositiveSample, Negative Sample, and Bad Sample. By introducing the Bad Sample (i.e. target sample) category, the purpose of suppressing Bad Samples can be achieved.

[0111] In one example, such as Figure 2d As shown, the fine-sorting stage can adopt the following steps:

[0112] 1. The target sample, positive samples and negative samples in the training sample can be fused to obtain samples of 3 categories (i.e., training sample set).

[0113] 2. It can acquire user features (such as user interest-based embeddings, content-based embeddings) and song features (such as lyric features, audio features) to train deep neural networks (e.g. Figure 3 As shown in the figure, the concatenated feature vector corresponding to any sample in the training sample set is input into the song re-ranking model;

[0114] 3. Obtain the model's predicted scores for each of the three categories of samples;

[0115] 4. Similar to the coarse-sorting stage, when calculating the loss, you can first go through SoftMax and then calculate the Cross Entropy loss.

[0116] 5. Predict the probability that a sample belongs to the Positive Sample category, and use this as the final model score, i.e., the model output result, which is used to characterize the probability information that any sample belongs to the positive sample type.

[0117] In another example, by mixing Bad Samples with the original Positive and Negative Samples in the training samples, the Bad Samples can be learned to distinguish them during the fine-ranking stage, thereby suppressing Bad Samples to some extent.

[0118] In one alternative embodiment, the fine ranking stage can also further improve the model's learning ability through methods such as user interest modeling, long sequence modeling, cross-scenario learning, and multi-objective learning, without being limited to these methods.

[0119] In this embodiment, a training sample set for the song ranking model is generated based on the target sample, positive samples, and negative samples in the training sample. Then, the user features and song features of each sample in the training sample set are concatenated to obtain the concatenated feature vector corresponding to each sample in the training sample set. The concatenated feature vector corresponding to any sample in the training sample set is then input into the song ranking model to obtain the model output result. This can better balance the matching degree between users and the songs being promoted and the ranking target, thereby improving the matching degree between users and the songs being promoted.

[0120] In one embodiment, in step S103, the training samples are used as training data for the song matching model and the song coarse ranking model, and the target samples and training samples are used as training data for the song fine ranking model. Joint model training of the song matching model, the song coarse ranking model, and the song fine ranking model may include the following steps:

[0121] Obtain the first loss of the song matching model and the second loss of the song coarse ranking model when training the song matching model and the song coarse ranking model using training samples; obtain the third loss of the song fine ranking model when training the song fine ranking model using target samples and training samples; perform weighted fusion of the first loss, the second loss and the third loss, and use the obtained fusion loss to update the parameters of the song matching model, the song coarse ranking model and the song fine ranking model.

[0122] In one example, during joint model training, the losses from the matching, coarse-ranking, and fine-ranking stages (i.e., the first loss, the second loss, and the third loss) can be weighted and fused to jointly train the song matching model, the coarse-ranking model, and the fine-ranking model. During model application, only the fine-ranking model trained in the fine-ranking stage is used as the song distribution model. For example, based on the probability score of the Positive Sample output by the model, the songs with the highest probability scores can be selected for distribution.

[0123] For example, the loss (i.e., the first loss) can be calculated using Sampled Softmax in the following manner:

[0124]

[0125] Where τ is the temperature coefficient. The smaller the temperature coefficient, the more emphasis is placed on learning difficult samples, and the more uniform the resulting representation space is. When calculating the loss, the batch matrix can be transposed to better suit the business scenario of the deployment.

[0126] For example, the loss can be calculated using Softmax Loss in the following way: first, calculate the Softmax based on the scores output by the coarse-ranking model for different learning objectives, and then calculate the loss (i.e., the second and third losses):

[0127]

[0128]

[0129] Where C corresponds to three categories (Positive Sample, Negative Sample, Bad Sample), y c For the corresponding sample's category label, the f function is used for Softmax normalization, and the output f(z) is... c () represents the probability of the corresponding category.

[0130] In another example, by jointly training models across multiple stages, each stage can perform its specific function. This allows for a good balance between the matching score and the ranking score of users and the songs to be delivered, based on the matching and ranking stages. It can also balance computational efficiency based on the coarse ranking and sampling stages. Thus, by adopting an End2End approach, suitable songs can be delivered to suitable user groups more effectively. While maximizing delivery efficiency, it also ensures compatibility between users and songs. Compared to traditional coarse-grained delivery methods, this approach significantly improves song delivery efficiency.

[0131] In this embodiment, by obtaining the first loss of the song matching model and the second loss of the song coarse ranking model when training the song matching model and the song coarse ranking model with training samples, and then obtaining the third loss of the song fine ranking model when training the song fine ranking model with target samples and training samples, the first loss, the second loss, and the third loss are weighted and fused. The parameters of the song matching model, the song coarse ranking model, and the song fine ranking model are updated using the obtained fused loss. This can improve the matching degree between users and the songs being delivered, while also improving the efficiency of song delivery.

[0132] In one embodiment, such as Figure 4 As shown, a song delivery method is provided. This embodiment illustrates the method applied to a server, but it is understood that the method can also be applied to a terminal, or to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps S401 to S402. Wherein:

[0133] In step S401, when it is detected that the user to be targeted meets the song targeting trigger conditions, the target song is determined from multiple candidate songs using the song targeting model; wherein, the song targeting model is trained by any of the song targeting model training methods.

[0134] In step S402, the target song is sent to the users to be targeted.

[0135] The song delivery model is a trained fine-ranking model. The trained fine-ranking model is obtained by using training samples as training data for the song matching model and the song coarse-ranking model, and target samples and training samples as training data for the fine-ranking model. The song matching model, the song coarse-ranking model and the fine-ranking model to be trained are jointly trained. The target samples are sampled from the first sample selected by the song matching model and the second sample selected by the song coarse-ranking model.

[0136] In practical applications, when the model is used for prediction, if it is detected that the user to be delivered meets the song delivery trigger conditions, such as when a user request is received, the probability that (User, Item) is a Positive Sample can be estimated for the current User (i.e., the user to be delivered) and the candidate Item (i.e. the candidate song to be delivered). This probability is used as the model score, and then the Item with the highest model score (i.e. the target song to be delivered) can be selected to obtain the final song that can be delivered.

[0137] In the above-mentioned song delivery method, when the user to be delivered meets the song delivery triggering conditions, the song delivery model determines the target song from multiple candidate songs and then delivers the target song to the user to be delivered. This method can better balance the matching degree and ranking target between the user and the song, improve the matching degree between the user and the song, and also improve the efficiency of song delivery.

[0138] In one embodiment, such as Figure 5 The diagram illustrates a training method for another song distribution model. In this embodiment, the method includes the following steps:

[0139] In step 501, multiple constructed samples are obtained based on multiple sample users and multiple sample songs in the training samples. A song matching model is used to obtain the matching prediction results of each constructed sample, and the first sample is selected from the multiple constructed samples. In step 502, a song coarse ranking model is used to obtain the ranking prediction results of each constructed sample, and the second sample is selected from the multiple constructed samples. In step 503, multiple candidate samples are obtained based on the intersection of the first sample and the second sample. The sampling quantity and preset sampling information are determined. The multiple candidate samples are sampled according to the sampling quantity and preset sampling information to obtain the target sample. In step 504, a training sample set for the song fine ranking model is generated based on the target sample, positive samples, and negative samples in the training samples. The user features and song features of each sample in the training sample set are concatenated to obtain the concatenated feature vector corresponding to each sample in the training sample set. In step 505, the concatenated feature vector corresponding to any sample in the training sample set is input into the song fine ranking model to obtain the model output result. In step 506, the first loss of the song matching model and the second loss of the song coarse ranking model are obtained when training the song matching model and the song coarse ranking model using training samples. The third loss of the song fine ranking model is also obtained when training the song fine ranking model using target samples and training samples. In step 507, the first, second, and third losses are weighted and fused. The resulting fused loss is used to update the parameters of the song matching model, the song coarse ranking model, and the song fine ranking model, resulting in the trained song fine ranking model. It should be noted that the specific limitations of the above steps can be found in the specific limitations of the training method for a song distribution model described above, and will not be repeated here.

[0140] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0141] Based on the same inventive concept, this application also provides a training device for a song distribution model to implement the training method for the song distribution model described above. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more training device embodiments for song distribution models provided below can be found in the limitations of the training method for song distribution models described above, and will not be repeated here.

[0142] In one embodiment, such as Figure 6 As shown, a training device for a song delivery model is provided, comprising:

[0143] The model acquisition module 601 is used to acquire training samples, a song matching model for predicting the matching degree between sample users and sample songs in the training samples, and a song coarse ranking model for coarsely ranking the pairing combinations of sample users and sample songs in the training samples.

[0144] The target sample obtaining module 602 is used to sample target samples from the first sample filtered by the song matching model based on the training sample and the second sample filtered by the song coarse ranking model based on the training sample; the matching degree between the sample user and the sample song in the target sample is less than the matching threshold, and the coarse ranking of the pairing combination of the sample user and the sample song in the target sample is greater than the ranking threshold.

[0145] The model joint training module 603 is used to use the training samples as training data for the song matching model and the song coarse ranking model, and to use the target samples and the training samples as training data for the song fine ranking model; to jointly train the song matching model, the song coarse ranking model and the song fine ranking model to obtain the trained song fine ranking model; the trained song fine ranking model is used as the song distribution model.

[0146] In one embodiment, the apparatus further includes:

[0147] The sample construction module is used to obtain multiple constructed samples based on multiple sample users and multiple sample songs in the training samples;

[0148] The first sample determination module is used to obtain the matching prediction results of each of the constructed samples using the song matching model, and to select the first sample from the plurality of constructed samples; the matching prediction results are used to characterize the degree of matching between the sample user and the sample song in the corresponding constructed sample;

[0149] The second sample determination module is used to obtain the ranking prediction results of each of the constructed samples using the song coarse ranking model, and to select the second sample from the plurality of constructed samples; the ranking prediction results are used to characterize the coarse ranking of the constructed samples that are paired and combined by sample users and sample songs.

[0150] In one embodiment, the first sample determination module includes:

[0151] The first feature input submodule is used to input the sample user features in each of the constructed samples into the user-side model of the song matching model, and to input the sample song features in each of the constructed samples into the song-side model of the song matching model.

[0152] The matching prediction result acquisition submodule is used to obtain the matching prediction result of each constructed sample based on the output result of the user-side model in the song matching model and the output result of the song-side model in the song matching model.

[0153] The first sample acquisition submodule is used to take the constructed sample corresponding to the matching prediction result that is less than the matching threshold as the first sample.

[0154] In one embodiment, the second sample determination module includes:

[0155] The second feature input submodule is used to input the sample user features in each of the constructed samples into the user-side model in the song coarse ranking model, and to input the sample song features in each of the constructed samples into the song-side model in the song coarse ranking model.

[0156] The ranking prediction result acquisition submodule is used to obtain the ranking prediction result of each constructed sample based on the target result output by the user-side model in the song coarse ranking model and the target result output by the song-side model in the song coarse ranking model; the target result is the output vector obtained by taking the type of the positive sample as the learning target;

[0157] The second sample acquisition submodule is used to construct samples corresponding to the ranking prediction results that are greater than the ranking threshold, and use them as the second sample.

[0158] In one embodiment, the target sample obtaining module 602 includes:

[0159] The candidate sample acquisition submodule is used to obtain multiple candidate samples based on the intersection of the first sample and the second sample;

[0160] The sampling information determination submodule is used to determine the number of samples and the preset sampling information; the preset sampling information is used to indicate the sampling method determined based on the matching prediction results and ranking prediction results of each candidate sample;

[0161] The sample sampling submodule is used to sample the multiple candidate samples according to the sampling quantity and the preset sampling information to obtain the target sample.

[0162] In one embodiment, the second sample determination module includes:

[0163] The training sample set generation module is used to generate a training sample set for the song ranking model based on the target sample, positive sample and negative sample in the training sample;

[0164] The feature vector concatenation module is used to concatenate the user features and song features of each sample in the training sample set to obtain the concatenated feature vectors corresponding to each sample in the training sample set.

[0165] The model output module is used to input the concatenated feature vector corresponding to any sample in the training sample set into the song ranking model to obtain the model output result; the model output result is used to characterize the probability information of any sample belonging to the type of the positive sample, and the probability information is used to determine whether to perform the song distribution operation.

[0166] In one embodiment, the model joint training module 603 includes:

[0167] The first loss and second loss determination submodule is used to obtain the first loss of the song matching model and the second loss of the song coarse ranking model when the training samples are used to train the song matching model and the song coarse ranking model.

[0168] The third loss determination submodule is used to obtain the third loss of the song re-ranking model when the song re-ranking model is trained using the target sample and the training sample;

[0169] The weighted fusion submodule is used to perform weighted fusion of the first loss, the second loss, and the third loss, and use the obtained fusion loss to update the parameters of the song matching model, the song coarse ranking model, and the song fine ranking model.

[0170] The modules in the training device for the aforementioned song delivery model can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0171] Based on the same inventive concept, this application also provides a song delivery device for implementing the song delivery method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations in one or more song delivery device embodiments provided below can be found in the limitations of the song delivery method described above, and will not be repeated here.

[0172] In one embodiment, such as Figure 7 As shown, a song delivery device is provided, comprising:

[0173] The song delivery model processing module 701 is used to determine the target song from multiple candidate songs by means of a song delivery model when it is detected that the user to be delivered meets the song delivery triggering conditions; wherein, the song delivery model is trained by any of the song delivery model training methods described above.

[0174] The target song delivery module 702 is used to deliver the target song to the user to be targeted.

[0175] Each module in the aforementioned song delivery device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0176] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 8 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores training data for a song delivery model. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a training method for a song delivery model.

[0177] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 9 As shown, this computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores song playback data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a song playback method.

[0178] Those skilled in the art will understand that Figure 8 and Figure 9 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0179] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0180] The system includes: acquiring training samples; a song matching model for predicting the matching degree between sample users and sample songs in the training samples; and a song ranking model for coarsely ranking the pairing combinations of sample users and sample songs in the training samples.

[0181] Target samples are sampled from the first sample selected by the song matching model based on the training samples and the second sample selected by the song coarse ranking model based on the training samples; the matching degree between the sample users and the sample songs in the target samples is less than the matching threshold, and the coarse ranking of the pairing combinations of sample users and sample songs in the target samples is greater than the ranking threshold.

[0182] The training samples are used as training data for the song matching model and the song coarse ranking model, and the target samples and the training samples are used as training data for the song fine ranking model; the song matching model, the song coarse ranking model and the song fine ranking model are jointly trained to obtain the trained song fine ranking model; the trained song fine ranking model is used as the song distribution model.

[0183] In one embodiment, the processor, when executing the computer program, also implements the steps of the training method for the song delivery model in the other embodiments described above.

[0184] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0185] When it is detected that the user to be targeted meets the song targeting trigger conditions, the target song is determined from multiple candidate songs by the song targeting model; wherein, the song targeting model is trained by the training method of any of the song targeting models described above;

[0186] The target song will be sent to the users to be targeted.

[0187] In one embodiment, the processor, when executing a computer program, also implements the steps of the song delivery method in the other embodiments described above.

[0188] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0189] The system includes: acquiring training samples; a song matching model for predicting the matching degree between sample users and sample songs in the training samples; and a song ranking model for coarsely ranking the pairing combinations of sample users and sample songs in the training samples.

[0190] Target samples are sampled from the first sample selected by the song matching model based on the training samples and the second sample selected by the song coarse ranking model based on the training samples; the matching degree between the sample users and the sample songs in the target samples is less than the matching threshold, and the coarse ranking of the pairing combinations of sample users and sample songs in the target samples is greater than the ranking threshold.

[0191] The training samples are used as training data for the song matching model and the song coarse ranking model, and the target samples and the training samples are used as training data for the song fine ranking model; the song matching model, the song coarse ranking model and the song fine ranking model are jointly trained to obtain the trained song fine ranking model; the trained song fine ranking model is used as the song distribution model.

[0192] In one embodiment, when the computer program is executed by a processor, it also implements the steps of the training method for the song delivery model in the other embodiments described above.

[0193] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0194] When it is detected that the user to be targeted meets the song targeting trigger conditions, the target song is determined from multiple candidate songs by the song targeting model; wherein, the song targeting model is trained by the training method of any of the song targeting models described above;

[0195] The target song will be sent to the users to be targeted.

[0196] In one embodiment, when the computer program is executed by a processor, it also implements the steps of the song delivery method in the other embodiments described above.

[0197] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, performs the following steps:

[0198] The system includes: acquiring training samples; a song matching model for predicting the matching degree between sample users and sample songs in the training samples; and a song ranking model for coarsely ranking the pairing combinations of sample users and sample songs in the training samples.

[0199] Target samples are sampled from the first sample selected by the song matching model based on the training samples and the second sample selected by the song coarse ranking model based on the training samples; the matching degree between the sample users and the sample songs in the target samples is less than the matching threshold, and the coarse ranking of the pairing combinations of sample users and sample songs in the target samples is greater than the ranking threshold.

[0200] The training samples are used as training data for the song matching model and the song coarse ranking model, and the target samples and the training samples are used as training data for the song fine ranking model; the song matching model, the song coarse ranking model and the song fine ranking model are jointly trained to obtain the trained song fine ranking model; the trained song fine ranking model is used as the song distribution model.

[0201] In one embodiment, when the computer program is executed by a processor, it also implements the steps of the training method for the song delivery model in the other embodiments described above.

[0202] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, performs the following steps:

[0203] When it is detected that the user to be targeted meets the song targeting trigger conditions, the target song is determined from multiple candidate songs by the song targeting model; wherein, the song targeting model is trained by the training method of any of the song targeting models described above;

[0204] The target song will be sent to the users to be targeted.

[0205] In one embodiment, when the computer program is executed by a processor, it also implements the steps of the song delivery method in the other embodiments described above.

[0206] 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, data stored, data displayed, 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 related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0207] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media 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), magnetic 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. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0208] The technical features of the above embodiments can be combined in any way. For the sake of brevity, 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.

[0209] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A training method for a song delivery model, characterized in that, The method includes: The system acquires training samples, a song matching model, and a song coarse-ranking model. The inputs to both the user-side model and the user-side model in the song matching model and the song coarse-ranking model are the user features of the sample users in the training samples. The inputs to both the song-side model and the song-side model in the song matching model are the song features of the sample songs in the training samples. The song matching model uses the inner product of the last layer output of the user-side model and the last layer output of the song-side model as the matching score. The matching score characterizes the degree of matching between the sample users and sample songs in the training samples. The song coarse-ranking model uses the inner product of the target result of the last layer output of the user-side model and the target result of the last layer output of the song-side model as the coarse-ranking score. The target result is the output vector obtained by taking the type of the positive sample as the learning objective. The coarse-ranking score characterizes the coarse-ranking of the pairings of sample users and sample songs in the corresponding training samples. Target samples are sampled from the first sample selected by the song matching model based on the training samples and the second sample selected by the song coarse ranking model based on the training samples; the matching degree between the sample users and the sample songs in the target samples is less than the matching threshold, and the coarse ranking of the pairing combinations of sample users and sample songs in the target samples is greater than the ranking threshold. The training samples are used as training data for the song matching model and the song coarse ranking model, and the target samples and the training samples are used as training data for the song fine ranking model; the song matching model, the song coarse ranking model and the song fine ranking model are jointly trained to obtain the trained song fine ranking model; the trained song fine ranking model is used as the song distribution model.

2. The method according to claim 1, characterized in that, Before the step of sampling the target sample from the first sample selected by the song matching model based on the training samples and the second sample selected by the song coarse ranking model based on the training samples, the method further includes: Based on multiple sample users and multiple sample songs in the training samples, multiple constructed samples are obtained; The matching prediction results of each constructed sample are obtained using the song matching model, and the first sample is selected from the plurality of constructed samples; the matching prediction results are used to characterize the degree of matching between the sample user and the sample song in the corresponding constructed sample; The ranking prediction results of each constructed sample are obtained using the song coarse ranking model, and the second sample is selected from the multiple constructed samples; the ranking prediction results are used to characterize the coarse ranking of the constructed samples that are paired and combined by sample users and sample songs.

3. The method according to claim 2, characterized in that, The step of obtaining matching prediction results for each constructed sample using the song matching model and selecting the first sample from the plurality of constructed samples includes: The sample user features in each of the constructed samples are input into the user-side model of the song matching model, and the sample song features in each of the constructed samples are input into the song-side model of the song matching model. Based on the output results of the user-side model and the song-side model in the song matching model, the matching prediction results of each constructed sample are obtained. The constructed sample corresponding to the matching prediction result that is less than the matching threshold is used as the first sample.

4. The method according to claim 2, characterized in that, The step of obtaining the ranking prediction results of each constructed sample using the song coarse ranking model, and selecting the second sample from the plurality of constructed samples, includes: The sample user features in each of the constructed samples are input into the user-side model of the song coarse ranking model, and the sample song features in each of the constructed samples are input into the song-side model of the song coarse ranking model. Based on the target results output by the user-side model and the target results output by the song-side model in the coarse-grained song ranking model, the ranking prediction results of each constructed sample are obtained; the target result is the output vector obtained by taking the type of the positive sample as the learning target. The constructed sample corresponding to the ranking prediction result that is greater than the ranking threshold is used as the second sample.

5. The method according to claim 1, characterized in that, The step of sampling target samples from the first sample selected by the song matching model based on the training samples and the second sample selected by the song coarse ranking model based on the training samples includes: Multiple candidate samples are obtained based on the intersection of the first sample and the second sample; The sampling quantity and preset sampling information are determined; the preset sampling information is used to indicate the sampling method determined based on the matching prediction results and ranking prediction results of each candidate sample. The target sample is obtained by sampling the multiple candidate samples according to the sampling quantity and the preset sampling information.

6. The method according to claim 1, characterized in that, The method further includes: Based on the target sample, the positive and negative samples in the training sample, a training sample set is generated for the song ranking model; The user features and song features of each sample in the training sample set are concatenated to obtain the concatenated feature vector corresponding to each sample in the training sample set. The concatenated feature vector corresponding to any sample in the training sample set is input into the song ranking model to obtain the model output result; the model output result is used to characterize the probability information that any sample belongs to the type of the positive sample, and the probability information is used to determine whether to perform the song distribution operation.

7. The method according to any one of claims 1 to 6, characterized in that, The step of using the training samples as training data for the song matching model and the song coarse ranking model, and the target samples and the training samples as training data for the song fine ranking model, to jointly train the song matching model, the song coarse ranking model, and the song fine ranking model includes: Obtain the first loss of the song matching model and the second loss of the song coarse ranking model when the training samples are used to train the song matching model and the song coarse ranking model; Obtain the third loss of the song re-ranking model when training the song re-ranking model using the target sample and the training sample; The first loss, the second loss, and the third loss are weighted and fused together, and the resulting fusion loss is used to update the parameters of the song matching model, the song coarse ranking model, and the song fine ranking model.

8. A method for delivering songs, characterized in that, The method includes: When it is detected that the user to be targeted meets the song targeting trigger conditions, the target song is determined from multiple candidate songs by the song targeting model; wherein, the song targeting model is trained by the training method of the song targeting model according to any one of claims 1-7; The target song will be sent to the users to be targeted.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7, and / or the steps of the method according to claim 8.

10. 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 steps of the method according to any one of claims 1 to 7, and / or the steps of the method according to claim 8.

Images