Song recommendation method, computer equipment and storage medium

Abstract

The invention relates to a song recommendation method. The method comprises the following steps: in response to a song recommendation request of a user, obtaining at least one to-be-recommended song corresponding to the song recommendation request; performing feature extraction processing of various different behavior targets on the user features of the user and the song features of each to-be-recommended song to obtain target features of each to-be-recommended song under each behavior target; wherein the behavior target is used for representing an interaction behavior of the expected user on the target recommended song; according to the target feature of each to-be-recommended song under each behavior target, the user feature and the song feature of each to-be-recommended song, sorting the at least one to-be-recommended song to obtain a sorting result of the at least one to-be-recommended song; and according to the sorting result of the at least one to-be-recommended song, determining a target recommended song corresponding to the song recommendation request from the at least one to-be-recommended song. By adopting the method, the song recommendation effect can be improved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a song recommendation method, apparatus, computer equipment, storage medium, and computer program product. Background Art

[0002] With the development of big data and artificial intelligence technologies, computer devices can more accurately capture user emotions and product preferences. In the music recommendation scenario, computer devices can provide users with personalized music recommendations, improving their listening experience.

[0003] Traditionally, song recommendations are often made through social networks and coordinated filtering. For example, song recommendations can be made based on the song preferences of other users with similar listening habits. Another example is the songs liked or frequently listened to by the user's social network friends. However, the songs preferred by social network friends or other users with similar listening habits may not necessarily be the user's own favorites, resulting in poor song recommendation results. Summary of the Invention

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

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

[0006] In response to a song recommendation request from a user, obtaining at least one song to be recommended corresponding to the song recommendation request;

[0007] Performing feature extraction processing for multiple different behavioral targets on the user features of the user and the song features of each of the songs to be recommended, to obtain target features of each of the songs to be recommended under each of the behavioral targets; wherein the behavioral target is used to characterize the interactive behavior that the user is expected to perform on the target recommended song;

[0008] Sorting the at least one song to be recommended according to the target feature of each song to be recommended under each behavior target, the user feature, and the song feature of each song to be recommended, to obtain a ranking result of the at least one song to be recommended;

[0009] According to the ranking result of the at least one song to be recommended, a target recommended song corresponding to the song recommendation request is determined from the at least one song to be recommended.

[0010] In a second aspect, the present application further provides a song recommendation device. The device comprises:

[0011] A song acquisition module, configured to respond to a user's song recommendation request and acquire at least one song to be recommended corresponding to the song recommendation request;

[0012] a feature extraction module configured to perform feature extraction processing for a plurality of different behavioral targets on the user features of the user and the song features of each of the songs to be recommended, to obtain target features of each of the songs to be recommended under each of the behavioral targets; wherein the behavioral targets are used to characterize the interactive behaviors that the user is expected to perform on the target recommended songs;

[0013] a song sorting module, configured to sort the at least one song to be recommended according to the target features of each song to be recommended under each of the behavioral targets, the user features, and the song features of each song to be recommended, to obtain a sorting result of the at least one song to be recommended;

[0014] The song recommendation module is used to determine the target recommended song corresponding to the song recommendation request from the at least one song to be recommended according to the sorting result of the at least one song to be recommended.

[0015] In a third aspect, the present application further provides a computer device. The computer device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are performed:

[0016] In response to a song recommendation request from a user, obtaining at least one song to be recommended corresponding to the song recommendation request;

[0017] Performing feature extraction processing for multiple different behavioral targets on the user features of the user and the song features of each of the songs to be recommended, to obtain target features of each of the songs to be recommended under each of the behavioral targets; wherein the behavioral target is used to characterize the interactive behavior that the user is expected to perform on the target recommended song;

[0018] Sorting the at least one song to be recommended according to the target feature of each song to be recommended under each behavior target, the user feature, and the song feature of each song to be recommended, to obtain a ranking result of the at least one song to be recommended;

[0019] According to the ranking result of the at least one song to be recommended, a target recommended song corresponding to the song recommendation request is determined from the at least one song to be recommended.

[0020] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the following steps:

[0021] In response to a song recommendation request from a user, obtaining at least one song to be recommended corresponding to the song recommendation request;

[0022] Performing feature extraction processing for multiple different behavioral targets on the user features of the user and the song features of each of the songs to be recommended, to obtain target features of each of the songs to be recommended under each of the behavioral targets; wherein the behavioral target is used to characterize the interactive behavior that the user is expected to perform on the target recommended song;

[0023] Sorting the at least one song to be recommended according to the target feature of each song to be recommended under each behavior target, the user feature, and the song feature of each song to be recommended, to obtain a ranking result of the at least one song to be recommended;

[0024] According to the ranking result of the at least one song to be recommended, a target recommended song corresponding to the song recommendation request is determined from the at least one song to be recommended.

[0025] In a fifth aspect, the present application further provides a computer program product. The computer program product includes a computer program that, when executed by a processor, implements the following steps:

[0026] In response to a song recommendation request from a user, obtaining at least one song to be recommended corresponding to the song recommendation request;

[0027] Performing feature extraction processing for multiple different behavioral targets on the user features of the user and the song features of each of the songs to be recommended, to obtain target features of each of the songs to be recommended under each of the behavioral targets; wherein the behavioral target is used to characterize the interactive behavior that the user is expected to perform on the target recommended song;

[0028] Sorting the at least one song to be recommended according to the target feature of each song to be recommended under each behavior target, the user feature, and the song feature of each song to be recommended, to obtain a ranking result of the at least one song to be recommended;

[0029] According to the ranking result of the at least one song to be recommended, a target recommended song corresponding to the song recommendation request is determined from the at least one song to be recommended.

[0030] The above-mentioned song recommendation method, device, computer equipment, storage medium and computer program product first obtain the songs to be recommended related to the user's song recommendation request, and then through feature extraction processing of multiple different behavioral targets, capture the target features under each behavioral target that meet the expectations from the user features and the song features of the songs to be recommended, so as to more comprehensively and accurately analyze the user's interest in the recommended songs, and then use the extracted target features, user features and song features under each behavioral target to sort the songs to be recommended, so that the target recommended songs obtained are more in line with the user's preferences, thereby achieving personalized and accurate song recommendations, and also increasing the probability of users interacting with the target recommended songs, thereby effectively improving the song recommendation effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 1. A schematic flow chart of a song recommendation method according to an embodiment;

[0032] Figure 2 A flowchart illustrating steps of performing multi-objective feature extraction processing on user features of a user and song features of songs to be recommended in one embodiment;

[0033] Figure 3 Schematic diagram of the model structure of a multi-objective processing model in one embodiment;

[0034] Figure 4 1 is a flow chart of steps for performing multi-target feature extraction processing on the first fusion feature in one embodiment;

[0035] Figure 5 Schematic diagram of the principle of the target attention mechanism in one embodiment;

[0036] Figure 6 A flowchart of a song recommendation method according to another embodiment;

[0037] Figure 7 1 is a flow chart of a song recommendation method according to another embodiment;

[0038] Figure 8 is a structural block diagram of a song recommendation device in one embodiment;

[0039] Figure 9 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0041] 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 regulations.

[0042] In one embodiment, Figure 1 As shown, a song recommendation method is provided. This embodiment uses the method applied to a terminal as an example for illustration. 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 implemented through the interaction between the terminal and the server. The terminal can be deployed with 8 cores of a Linux CPU, and can also be deployed with a deep learning framework to further improve the performance and efficiency of the terminal in executing the song recommendation method; the deep learning framework can be implemented through TensorFlow (an open source machine learning framework) or through PyTorch (an open source deep learning framework for machine learning and deep learning). In this embodiment, the method includes the following steps S101-S104:

[0043] Step S101: In response to a song recommendation request from a user, obtaining at least one song to be recommended corresponding to the song recommendation request.

[0044] Here, a user refers to an account in a music scenario, for example, an account logged in on a music application, a music platform, or a music mini-program.

[0045] The song recommendation request refers to a message instruction requesting the recommendation of songs that the user may be interested in. For example, the song recommendation request may be a request to recommend songs of a specific genre, singer, or instrument.

[0046] The songs to be recommended are multiple recommended songs obtained through preliminary song recommendation processing. For example, the songs to be recommended may be multiple recommended songs obtained through social network recommendation, collaborative filtering, content recommendation, etc.

[0047] Specifically, the terminal may pre-install a music program in the form of an application (APP), mini-program, or other similar format. After logging into the music program, the user can initiate a song recommendation request through the music program. For example, the user enters a song recommendation request through the music program (e.g., via an input box or voice input). The terminal then makes a song recommendation based on the song recommendation request, such as initially processing it through content recommendation, deep learning models, or semantic analysis to obtain at least one recommended song, and recommend songs that match the input content.

[0048] Alternatively, the user triggers a page button of the music program to make an automatic recommendation, for example, through social network recommendation, collaborative filtering and other preliminary processing methods, at least one song that the user or the user's social friends likes is obtained as a song to be recommended.

[0049] For example, if a user enters a song title, singer name, or album name in the input box on the page provided by the music program, the terminal can then obtain songs that match the content data entered by the user and recommend songs. For example, it can retrieve N songs that match the content data entered by the user and set these N songs as songs to be recommended. To give another example, a recommendation request button is displayed in the interface provided by the music program. When the user triggers the recommendation request button, a song recommendation request is generated. The terminal needs to "guess" the songs that the user may like. After the user's authorization, the terminal can obtain N songs that the user's friends like through formal and legal channels and set these N songs as songs to be recommended.

[0050] Step S102, performing feature extraction processing for multiple different behavioral targets on the user features of the user and the song features of each song to be recommended, and obtaining target features of each song to be recommended under each behavioral target; wherein the behavioral target is used to characterize the interactive behavior expected of the user on the target recommended song.

[0051] User characteristics refer to important information that describes the various attributes and characteristics of individual users. These attributes and characteristics include, but are not limited to, demographic attributes, behavioral characteristics, preferences, and emotional tendencies.

[0052] Song features are important information that describes the various attributes and characteristics of a song. These include, but are not limited to, the song's style, emotion, musical structure, and audio characteristics.

[0053] In the music scenario, after recommending music to users, the target behavior is the interactive behavior that the user is expected to trigger in response to the recommended music. Interactive behaviors include but are not limited to likes, comments, forwarding, playing, and favorites. Target features are feature information related to the interactive behavior that the user is expected to achieve. For example, target features can be user information and song information related to the interactive behavior.

[0054] Specifically, the terminal can collect information such as the user's demographic attributes, characteristics of interactive behavior when listening to songs, and other information to extract the user's user characteristics. The terminal can also perform feature extraction processing on each song to be recommended to obtain the song characteristics of each song to be recommended. The terminal can pre-set the behavioral goals that the user is expected to achieve for the recommended songs, such as increasing the number of plays, comments, and likes of the songs; then, the terminal performs feature extraction processing on the user characteristics and song characteristics according to the multiple behavioral goals expected to be achieved, so as to obtain the target characteristics of each song to be recommended under each behavioral goal.

[0055] Step S103, sorting the at least one song to be recommended according to the target features of each song to be recommended under each behavior target, the user features and the song features of each song to be recommended, to obtain a sorting result of the at least one song to be recommended.

[0056] The ranking result represents the final recommendation result (or search result) obtained after ranking and scoring the initially obtained candidate data (e.g., songs to be recommended). In music recommendation, songs to be recommended that match the content data input by the user can be initially processed, or songs to be recommended can be initially obtained based on the user's or their social friends' favorite songs, and then these songs to be recommended can be ranked. It should be noted that after obtaining the songs to be recommended, the songs to be recommended can be initially ranked and then re-ranked in step S103; alternatively, the initial ranking can be omitted and ranking can be performed directly in step S103.

[0057] Re-ranking is a technique used in information retrieval, recommendation systems, or search engines. Re-ranking aims to further optimize the initially generated candidate list to improve the relevance of the final selection and user satisfaction.

[0058] Specifically, the terminal can adopt ranking learning algorithms, deep learning models, etc. to score at least one song to be recommended based on the target characteristics, user characteristics and song characteristics of each song to be recommended under each behavioral goal, so as to reflect the user's possible preference for each song to be recommended through the score, and then the output score information can be used as the ranking result of at least one song to be recommended.

[0059] Step S104: determining a target recommended song corresponding to the song recommendation request from the at least one song to be recommended according to the ranking result of the at least one song to be recommended.

[0060] The target recommended songs refer to the songs that are finally determined to be recommended to the user.

[0061] Specifically, the terminal may determine a target recommended song corresponding to the song recommendation request from at least one song to be recommended based on the sorting result represented by the rating information. For example, the terminal may select a song to be recommended whose rating information exceeds a score threshold as the target recommended song, or may select the M songs with the highest ratings in descending order as the target recommended songs. The terminal may display the target recommended song to the user so that the user can trigger an interaction with the target recommended song.

[0062] For example, in response to a user's request for song recommendation, the terminal uses a song recommendation algorithm to perform preliminary processing to obtain the Top N songs, and then uses the song features of these Top N songs and the user features of the user to perform feature extraction and sorting processing on these Top N songs for multiple different behavioral goals, and outputs the scores of these Top N songs; finally, based on the scoring results, the final list of recommended songs is determined from these Top N songs.

[0063] In the above song recommendation method, the songs to be recommended related to the user's song recommendation request are first obtained, and then through feature extraction processing of multiple different behavioral targets, the target features under each behavioral target that meet the expectations are captured from the user features and the song features of the songs to be recommended, so as to more comprehensively and accurately analyze the user's interest in the recommended songs, and then use the extracted target features, user features and song features under each behavioral target to sort the songs to be recommended, so that the target recommended songs obtained are more in line with the user's preferences, realizing personalized and accurate song recommendations, and also increasing the probability of users interacting with the target recommended songs, which effectively improves the song recommendation effect.

[0064] In one embodiment, Figure 2 As shown, in step S102, the user characteristics of the user and the song characteristics of each song to be recommended are subjected to feature extraction processing for multiple different behavioral targets to obtain the target features of each song to be recommended under each behavioral target, which specifically includes the following contents:

[0065] Step S201 : performing feature fusion processing on the user's user features and the song features of each song to be recommended to obtain a first fusion feature.

[0066] Specifically, the terminal splices the user's user features and the song features of the songs to be recommended. For example, the terminal may splice the song features of any song to be recommended at the end of the user features, or may splice the user features at the end of each song to be recommended, thereby obtaining the first fusion feature.

[0067] For example, assuming that the songs to be recommended include song 1, song 2, ..., song N, the song features of the songs to be recommended include song 1 features, song 2 features, ..., song N features. The terminal concatenates user A's user A features with each song feature, resulting in a first fused feature consisting of user A features + song 1 features, user A features + song 2 features, ..., user A features + song N features.

[0068] Step S202: Through the trained multi-objective processing model, the first fusion feature corresponding to each song to be recommended is subjected to feature extraction processing for multiple different behavioral targets to obtain the target features of each song to be recommended under each behavioral target; wherein the multi-objective processing model includes at least multiple expert networks, a gating network corresponding to each behavioral target, and a target task tower network corresponding to each behavioral target.

[0069] The multi-objective processing model is a feature extraction model that balances and optimizes multiple behavioral objectives. It is used in scenarios where multiple conflicting behavioral objectives must be considered simultaneously, extracting feature information for each objective. In practical applications, multi-objective processing models can be constructed based on multi-gate mixture of experts (MMOE) networks.

[0070] It's important to note that multi-objective modeling is a modeling requirement. For example, in a music recommendation scenario, after recommending a song to a user, you want them to click on it, listen to it for as long as possible, and receive frequent likes, comments, and reposts. Therefore, you need to simultaneously model four behavioral objectives: click-through rate, viewing time, like rate, comment rate, and repost rate.

[0071] Specifically, the terminal can first determine the multiple behavioral goals that it wants to improve the user's triggering of target recommended songs, and then improve the model structure of the multi-gated expert network based on the multiple behavioral goals to obtain a multi-target processing model to be trained; for example, construct a target task tower network corresponding to each behavioral goal; based on the gating network corresponding to each behavioral goal, multiple expert networks and the target task tower network corresponding to each behavioral goal, construct a multi-target processing model to be trained; use song feature samples and user feature samples to train the multi-target processing model to be trained to obtain a trained multi-target processing model.

[0072] The target task tower network (Tower) maps input data to a specific target output, such as classification, regression, or prediction. The target task tower network is connected to the expert network through a gating mechanism. The gating mechanism dynamically selects different expert networks based on the input. The target task tower uses the fused features of these selected expert networks to generate the final target features. This mechanism enables each task to utilize expert knowledge tailored to its specific needs.

[0073] Among them, the expert network (Expert) is used to learn and extract specific feature information of the input data (such as the first fusion feature). The multi-objective processing model can include multiple expert networks, each of which can be regarded as an expert. Different expert networks can focus on processing different feature information of the input data. Each expert network operates independently, effectively improving the computational efficiency and data processing speed of the multi-objective processing model. For example, suppose the multi-objective processing model contains two expert networks, one of which is used to extract feature vectors related to user features from the input data, and the other expert network can extract feature vectors related to song features from the input data.

[0074] Among them, the gating network (Gate) is used to adaptively and dynamically assign weights to each expert network based on the input data (such as the first fusion feature). On the one hand, the weights can be used to determine which expert networks can participate in the final output calculation. On the other hand, the weights can also be used to represent the importance of the expert network to the behavioral goal.

[0075] Figure 3 The schematic diagram of the model structure of the multi-objective processing model is shown in Figure 2. Figure 3As shown, the terminal can set corresponding behavioral targets for each task tower network as output targets. This set task tower network is called a target task tower network. It should be understood that there is a one-to-one correspondence between behavioral targets and target task tower networks. Furthermore, there is also a one-to-one correspondence between gating networks and behavioral targets. The terminal constructs a multi-target processing model to be trained using multiple expert networks, gating networks corresponding to each behavioral target, target task tower networks corresponding to each behavioral target, an input layer, a fusion layer, and an output layer. The input layer receives raw input data (such as the first fused feature) and can also preprocess the raw input data to adapt it to subsequent networks. For example, it can format the first fused feature into a format that can be processed by the expert network. The fusion layer weights and fuses the outputs of multiple experts to generate intermediate representation features (such as fused features). The output layer converts the output of the target task tower network into the desired prediction result, such as a regression value. The output layer typically uses different loss functions to process the outputs of each behavioral target. Subsequently, the song feature samples and user feature samples are used to iteratively train the multi-objective processing model to be trained, and finally a trained multi-objective processing model is obtained. Among them, the user feature samples can be obtained based on the historical information of historical users.

[0076] Furthermore, after processing and obtaining the song features and user features corresponding to each song to be recommended, the terminal can input the user features and song features into the trained multi-objective processing model, and output the target features of each song to be recommended under each behavioral goal.

[0077] In this embodiment, by splicing the user's user features and the song features of each song to be recommended into the first fusion feature, the trained multi-objective processing model performs more comprehensive feature extraction processing on each first fusion feature, which helps to improve the feature extraction effect of the trained multi-objective processing model under each behavioral target, thereby effectively improving the data quality of the target features obtained under each behavioral target, and providing a reliable data basis for sorting in subsequent steps.

[0078] In one embodiment, Figure 4 As shown, the above step S202 (using the trained multi-objective processing model to perform feature extraction processing for multiple different behavioral objectives on the first fusion feature corresponding to each song to be recommended, to obtain the target feature of each song to be recommended under each behavioral objective) specifically includes the following steps:

[0079] Step S401: input the first fusion feature into each expert network in the multiple expert networks respectively to obtain a feature vector output by each expert network.

[0080] like Figure 3 As shown, the terminal can first input the first fusion feature into the input layer, so as to convert the format of the first fusion feature into a format that can be processed by the expert network through the input layer, and obtain the preprocessed first fusion feature, and then input the preprocessed first fusion feature into each expert network in the multiple expert networks respectively, so as to process the preprocessed first fusion feature according to the corresponding behavioral target task through each expert network, so that each expert network learns the feature vectors under different behavioral targets as the feature vectors output by each expert network.

[0081] Taking the trained multi-objective processing model containing three expert networks as an example, the expert network Expert0 can be responsible for learning and extracting feature information of demographic attributes, the expert network Expert1 can be responsible for learning and extracting feature information of behavior, and the expert network Expert2 can be responsible for learning, extracting and processing feature information of songs. Then, the terminal can use the expert networks Expert0, Expert1 and Expert2 to perform different processing on the first fusion feature after preprocessing, thereby outputting a feature vector representing multiple aspects of information.

[0082] In step S402, the first fused features are input into the gating network corresponding to each behavioral target to obtain the weight of each expert network under each behavioral target. It should be noted that the above steps S401 and S402 can be performed simultaneously, or they can be performed sequentially.

[0083] Among them, the weight is used to measure the importance of the expert network in the behavioral goal.

[0084] Specifically, the multi-objective processing model can include the same number of gating networks as the set behavioral objectives. If there is only one gating network, the terminal inputs the first fusion feature into the gating network, and the gating network assigns weights to each expert network based on the first fusion feature to obtain a weight set, which contains the weights of each expert network; this weight set is input into the fusion layer. If there are multiple gating networks, the terminal can input the first fusion feature into each gating network respectively, and each gating network assigns weights to each expert network based on the first fusion feature to obtain a weight set output by each gating network. The fusion layer can then receive the weight set output by each gating network.

[0085] In practical applications, the number of gated networks can be related to the number of behavioral goals, and the number of gated networks can also be the same as the number of target task tower networks, so as to measure the importance of each expert network for each behavioral goal. Figure 3 As shown in the figure, it is assumed that the multi-objective processing model includes a gated network Gate A (corresponding to behavioral objective A) and a gated network Gate B (corresponding to behavioral objective B), and expert networks Expert0, Expert1, and Expert2. Based on the first fusion feature, the gated network Gate A performs weight allocation processing on the expert networks Expert0, Expert1, and Expert2 to obtain a weight set {Weight0, Weight1, Weight2}, where Weight0 represents the weight of the expert network Expert0 output by the gated network Gate A under behavioral objective A; Weight1 represents the weight of the expert network Expert1 output by the gated network Gate A under behavioral objective A; and Weight2 represents the weight of the expert network Expert2 output by the gated network Gate A under behavioral objective A. Based on the first fusion feature, gated network Gate B performs weight assignment on expert networks Expert0, Expert1, and Expert2, obtaining a weight set {Weight3, Weight4, Weight5}, where Weight3 represents the weight of expert network Expert0 under behavioral target B, as output by gated network Gate B; Weight4 represents the weight of expert network Expert1 under behavioral target B, as output by gated network Gate B; and Weight5 represents the weight of expert network Expert2 under behavioral target B, as output by gated network Gate B. The terminal inputs the weight sets {Weight0, Weight1, Weight2} and {Weight3, Weight4, Weight5} into the fusion layer.

[0086] Step S403 , performing weighted summation processing on the weight of each expert network under each behavior target and the feature vector output by each expert network, to obtain the fusion features of the song to be recommended under each behavior target.

[0087] Specifically, the fusion layer obtains the weight of each expert network under each behavioral goal from the weight set, and then performs weighted summation on the weight of each expert network under each behavioral goal and the feature vector output by each expert network to obtain the fusion features of the song to be recommended under each behavioral goal; then the fusion features of the song to be recommended under each behavioral goal are input into the target task tower network of the corresponding behavioral goal, that is, the knowledge of each expert network can be shared between the target task tower networks.

[0088] Step S404: Through the target task tower network corresponding to each behavior target, the fusion features of the song to be recommended under each behavior target are regressed to obtain the target features of the song to be recommended corresponding to each behavior target.

[0089] Specifically, the terminal can perform regression processing of different behavioral targets on the fusion features through the target task tower network corresponding to each behavioral target, and then the terminal obtains the target features output by each target task tower network, and then the terminal obtains the target features corresponding to each behavioral target of the song to be recommended. For example, the target task tower network can be Figure 3 As shown in Tower A and Tower B, the terminal can set "play songs" and "collect songs" as two behavioral goals. Then the target task tower network Tower A regresses the fusion features with "play songs" as the behavioral goal and outputs the target features corresponding to "play songs". The target task tower network Tower B regresses the fusion features with "collect songs" as the behavioral goal and outputs the target features corresponding to "collect songs". Then the target features corresponding to the behavioral goal of "playing songs" and the target features corresponding to the behavioral goal of "collecting songs" are the target features of the songs to be recommended under each behavioral goal respectively.

[0090] In this embodiment, the first fusion feature is input into each expert network respectively, so that each expert network can learn different feature information, and then obtain the feature vector output by each expert network, which effectively improves the diversity of the feature vector; the weight of each expert network under each behavioral target is calculated based on the input first fusion feature through the gating network, so that the importance of each expert network to the current behavioral target is reflected through the weight, ensuring that the multi-target processing model can dynamically select the most relevant expert network for output; the feature vector output by each expert network is weighted and summed to fuse the feature vectors of different expert networks, so that the target task tower network can have a more comprehensive understanding of the feature information; finally, the fusion feature of each behavioral target is regressed through the target task tower network of each behavioral target, so as to generate the target feature corresponding to each behavioral target, which not only improves the flexibility and adaptability of the multi-target processing model, but also enables the target task tower networks corresponding to different behavioral targets to effectively share the knowledge of the expert network, thereby enhancing the overall performance accuracy of the multi-target processing model.

[0091] In one embodiment, before the above-mentioned step S102 (performing feature extraction processing for multiple different behavioral targets on the user's user features and the song features of each song to be recommended, and obtaining the target features of each song to be recommended under each behavioral target), it also includes: based on the attention mechanism, performing feature extraction processing on the user's song sequence to obtain the user's interactive behavior sequence features; and obtaining the user's user features based on the user's user attribute features and the interactive behavior sequence features.

[0092] The user attribute characteristics refer to characteristics that describe the user's demographic attributes. For example, the user attribute characteristics may include demographic attributes such as gender, age, and education level.

[0093] The interactive behavior sequence feature is a feature that describes the interactive behavior between the user and the song. For example, the interactive behavior sequence feature can be a feature that describes the interactive behavior of the user when listening to the song, such as liking, collecting, and commenting.

[0094] Specifically, the terminal can use the attention mechanism to explore the differences in the impact of different songs on the user's current listening interests, thereby improving the ability of the obtained user interaction behavior sequence features to represent user interests. In turn, the user attribute features and interaction behavior sequence features can be set as the user's user features.

[0095] In this embodiment, the attention mechanism is used to perform feature extraction processing on the user's song sequence, which can strengthen relevant interests and weaken irrelevant interests, thereby improving the interest representation ability of obtaining the user's interactive behavior sequence characteristics; and then setting the user attribute characteristics and interactive behavior sequence characteristics as the user's user characteristics, thereby improving the diversity of the user feature representation ability, which helps to improve the feature processing effect of the multi-target processing model in the subsequent steps.

[0096] In one embodiment, the song sequence includes the user's historical play recommendation songs and the user's historical interaction songs; the historical play recommendation songs represent at least one song to be recommended that was played by the user within a historical time period; the historical interaction songs represent songs associated with the user's interaction behaviors under various different behavioral goals within the historical time period. The step of "performing feature extraction processing on the user's song sequence based on an attention mechanism to obtain the user's interaction behavior sequence features" specifically includes the following: performing attention mechanism processing on the historical play recommendation songs and historical interaction songs to obtain the user's behavior sequence features; and performing pooling processing on the behavior sequence features to obtain the user's interaction behavior sequence features.

[0097] Among them, the attention mechanism can be Target Attention, Figure 5 Schematic diagram of the target attention mechanism.

[0098] Specifically, if Figure 5 As shown, the terminal can determine the sequence of songs played by the user within a historical time period from the songs to be recommended. The terminal then obtains a historically played recommended song Q. The terminal can also separately obtain song sequences associated with the user's interactive behaviors under various behavioral objectives during the historical time period, such as the sequence of songs the user has liked and the sequence of songs the user has favorited. The terminal then obtains a historically played song KV consisting of the song sequences associated with each behavioral objective. Operations are performed on the historically played recommended song Q and the historically played song KV to obtain user sequence features. The user sequence features are then multiplied by the position encoding to obtain the user's behavioral sequence features. The terminal can also perform pooling on the behavioral sequence features. For example, sum-pooling can be performed on the behavioral sequence features. This involves dividing the behavioral sequence features into several regions, summing the feature information for each region, and finally concatenating the summed results from all regions as the output. Ultimately, the terminal obtains the user's interactive behavior sequence features.

[0099] In this embodiment, the user's historical play recommendation songs and historical interaction songs are processed through the target attention mechanism to enhance the feature information of the songs played and interacted by the user in the historical time period, thereby improving the feature expression ability of the processed interaction behavior sequence features for the songs that the user is interested in.

[0100] In one embodiment, the song sequence includes all songs played in the user's history; all songs played in the history represent all songs played by the user within a historical time period; and all songs played in the history include the user's historically recommended songs. Based on the attention mechanism, feature extraction is performed on the user's song sequence to obtain the user's interactive behavior sequence features, specifically including the following: feature extraction is performed on all songs played in the history to obtain historical song features of all songs played in the history; based on the similarity between the song features of each song to be recommended and the historical song features, an attention weight is obtained for each song to be recommended; based on the attention weight of each song to be recommended, a weighted summation is performed on at least one song to be recommended to obtain the user's interactive behavior sequence features.

[0101] Among them, the attention mechanism can also be Din (Deep Interest Network).

[0102] Specifically, the terminal obtains all songs played by the user within a historical time period, including songs that have been played among the songs to be recommended (i.e., historically played recommended songs) as well as songs that are not among the songs to be recommended but have been played. The terminal then extracts the overall features of all the historically played songs to obtain historically played song features. The terminal then determines the attention weight for each song to be recommended based on the similarity between the song features of each song to be recommended and the features of the historically played songs. Based on this similarity, the terminal then performs a weighted summation of the song features of all the songs to be recommended based on the attention weight of each song to be recommended to calculate the user's interaction behavior sequence features.

[0103] In this embodiment, a deep interest network is used to calculate the similarity between all songs played in the user's history and each song to be recommended, and the weight of each song to be recommended is determined based on the similarity, so as to strengthen the feature information of the songs to be recommended that are more similar to the songs played in the user's history, thereby improving the ability of the processed interactive behavior sequence features to express the features of the songs that the user is interested in.

[0104] In one embodiment, the above-mentioned step S103 (sorting at least one song to be recommended according to the target features, user features and song features of each song to be recommended under each behavioral goal, and obtaining a sorting result of at least one song to be recommended) specifically includes the following contents: performing feature fusion processing on the target features, user features and song features of each song to be recommended under each behavioral goal, and obtaining a second fused feature of each song to be recommended; sorting the second fused feature through a pre-trained sorting model, and obtaining a sorting result of at least one song to be recommended.

[0105] Ranking models are machine learning models used to optimize the sorting or ranking of songs to be recommended. In practical applications, ranking models use inputs such as features, user behavior, and contextual information based on initial candidates (e.g., songs to be recommended) to score or rank the songs, generating results that better meet user needs. In practical applications, ranking models can be local ranking functions.

[0106] Specifically, the terminal can also splice the target features, user features, and song features of each song to be recommended under each behavioral target, by splicing the target features of each song to be recommended under each behavioral target at the end of the first fusion feature, or by splicing the first fusion feature at the end of the target features of each song to be recommended under each behavioral target, and then the terminal processes and obtains the second fusion feature of each song to be recommended. The terminal inputs the second fusion feature into a pre-trained ranking model to score each song to be recommended based on the second fusion feature through the pre-trained ranking model, and outputs the scoring result as the ranking result of the song to be recommended.

[0107] In this embodiment, by splicing the target features, user features and song features of each song to be recommended under each behavioral target, the second fusion features of each song to be recommended can be made to reflect the feature information related to the user's interests more richly and comprehensively, further improving the quality of the feature data, providing reliable data support for the sorting model in subsequent steps, and helping to improve the sorting effect of the sorting model.

[0108] In one embodiment, the step of "sorting the second fusion feature through a pre-trained sorting model to obtain a sorting result of at least one song to be recommended" specifically includes the following contents: extracting context information from the second fusion feature through a context information processing model in the sorting model to obtain contextual information of the second fusion feature; sorting the at least one song to be recommended according to the context information through a multi-objective sorting model in the sorting model to obtain a sorting result of at least one song to be recommended.

[0109] The context information processing model is used to extract contextual information from the input data (e.g., the second fused feature). This contextual information includes the hidden layer outputs O1, O2, ..., On, and the latent vector Sn. The context information processing model can be constructed based on a gated recurrent unit (GRU).

[0110] Among them, the multi-objective ranking model is used to score and / or rank the initial candidate items (such as songs to be recommended) to optimize the ranking of the initial candidate items.

[0111] Specifically, the terminal can now input the second fusion feature into the context information processing model in the sorting model to extract context information from the second fusion feature through the context information processing model; input the context information into the multi-objective sorting model in the sorting model, and the multi-objective sorting model scores at least one song to be recommended based on the context information to obtain a scoring result of at least one song to be recommended; the scoring result can be output as a sorting result, and the scoring result can be used to sort at least one song to be recommended to obtain a sorted song to be recommended, and then the sorted song to be recommended and its scoring result can be output as a sorting result.

[0112] In this embodiment, the second fusion feature is subjected to context information extraction processing through the context information processing model in the sorting model, which helps to mine context information related to the user's demographic attributes, interactive behavior, and song preferences; through the multi-objective sorting model in the sorting model, at least one song to be recommended is sorted in combination with the context information, and the sorting result of at least one song to be recommended not only takes into account the user's preferences and interactive behavior habits, but also integrates a wider range of context information, thereby effectively improving the personalization and strong correlation of the sorting results, so that recommendations to users not only meet user needs, but also help to achieve multiple behavioral goals that users are expected to perform on songs, thereby greatly improving the recommendation effect of songs.

[0113] In one embodiment, the above-mentioned step S101 (responding to the user's song recommendation request, obtaining at least one song to be recommended corresponding to the song recommendation request) specifically includes the following contents: obtaining user input data corresponding to the song recommendation request; determining at least one song that matches the user input data from the music library, and obtaining at least one song to be recommended corresponding to the song recommendation request.

[0114] The user input data refers to information used to describe the user's song recommendation needs.

[0115] Specifically, the user can initiate a song recommendation request through the terminal, and then the terminal can collect user input data corresponding to the song recommendation request. For example, the user can input input data in text form through the input box displayed by the terminal in the music program interface, or can input input data in voice form by speaking, and the terminal will obtain the user input data corresponding to the song recommendation request. The terminal can first perform data preprocessing on the user input data, such as converting the data format, noise reduction, semantic recognition, text optimization, etc., to more accurately understand the user's song recommendation needs and obtain preprocessed data. The terminal performs song recommendation processing based on the preprocessed data. For example, it can search for at least one song that matches the preprocessed data from the music library, and can also recommend at least one song with a high similarity based on the characteristics of the preprocessed data (such as rhythm, instrument, style, etc.) through a recommendation algorithm. The terminal then obtains at least one song to be recommended corresponding to the song recommendation request.

[0116] In this embodiment, the user input data corresponding to the user's song recommendation request is first obtained to accurately understand the user's song recommendation needs, and then song recommendation processing is performed on the user input data to preliminarily obtain at least one song to be recommended that meets the song recommendation request, laying the foundation for subsequent song recommendation processing.

[0117] In one embodiment, Figure 6 As shown, another song recommendation method is provided, which is described by taking the application of the method to a terminal as an example, and includes the following steps:

[0118] Step S601: In response to a song recommendation request from a user, obtaining at least one song to be recommended corresponding to the song recommendation request.

[0119] Step S602 : performing feature fusion processing on the user's user features and the song features of each song to be recommended, and obtaining a first fusion feature corresponding to each song to be recommended.

[0120] Step S603: Through the trained multi-objective processing model, the first fusion feature corresponding to each song to be recommended is subjected to feature extraction processing for multiple different behavioral objectives to obtain the target features of each song to be recommended under each behavioral objective.

[0121] Step S604 , performing feature fusion processing on the target features, user features and song features of each song to be recommended under each behavior target, to obtain a second fusion feature of each song to be recommended.

[0122] Step S605 : Perform context information extraction processing on the second fused feature through the context information processing model in the sorting model to obtain context information of the second fused feature.

[0123] Step S606: sorting the at least one song to be recommended according to the context information using the multi-objective sorting model in the sorting model to obtain a sorting result for the at least one song to be recommended.

[0124] Step S607: Determine a target recommended song corresponding to the song recommendation request from the at least one song to be recommended according to the ranking result of the at least one song to be recommended.

[0125] The above song recommendation method can achieve the following beneficial effects: first, the songs to be recommended related to the user's song recommendation request are obtained, and then through feature extraction processing of multiple different behavioral targets, the target features under each behavioral target that meet the expectations are captured from the user features and the song features of the songs to be recommended, so as to more comprehensively and accurately analyze the user's interest in the recommended songs, and then use the extracted target features, user features and song features under each behavioral target to sort the songs to be recommended, so that the target recommended songs obtained are more in line with the user's preferences, realizing personalized and accurate song recommendations, and also increasing the probability of users interacting with the target recommended songs, thereby effectively improving the song recommendation effect.

[0126] In order to more clearly illustrate the song recommendation method provided by the embodiment of the present disclosure, the song recommendation method is specifically described below using a specific embodiment. Figure 7 As shown, another song recommendation method is provided, which can be applied to a terminal, specifically including the following contents:

[0127] In a music recommendation scenario, the terminal retrieves N songs based on a user's request. User features are obtained, including user attribute features and interaction behavior sequence features. User attribute features include demographic information such as gender, age, and education level. User listening sequence features are mined from the user's current song sequence using an attention mechanism. The song features and user features of the N songs are concatenated and used as input to a multi-gated expert network (MMOE). The target features output by the MMOE for each behavior goal (e.g., play and favorite) are then concatenated with the user features and song features for each song as input to a gated recurrent unit (GRU). The GRU learns information about the N songs and forms local context. The GRU's supervision labels are multiple goals (e.g., complete play and favorite). This results in a local context-focused ranking function that ranks the N songs based on this local context. Finally, the ranking result determines the final list of recommended songs.

[0128] In this embodiment, by taking playback behavior and collection behavior as multiple behavioral targets for feature extraction processing, the target features of multiple behavioral targets can be extracted more comprehensively, so as to utilize the target features of multiple behavioral targets to improve the sorting effect of songs, and significantly increase the number of times users play and collect the recommended song list.

[0129] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed 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 portion of steps or stages in other steps.

[0130] Based on the same inventive concept, embodiments of the present application also provide a song recommendation device for implementing the aforementioned song recommendation method. The solution provided by this device is similar to the solution described in the aforementioned method. Therefore, the specific limitations of one or more song recommendation device embodiments provided below can be found in the aforementioned limitations of the song recommendation method and will not be further elaborated here.

[0131] In one embodiment, Figure 8 As shown, a song recommendation device 800 is provided, comprising: a song acquisition module 801, a feature extraction module 802, a song sorting module 803 and a song recommendation module 804, wherein:

[0132] The song acquisition module 801 is used to respond to a song recommendation request from a user and acquire at least one song to be recommended corresponding to the song recommendation request.

[0133] The feature extraction module 802 is used to perform feature extraction processing for multiple different behavioral targets on the user features and the song features of each song to be recommended, and obtain the target features of each song to be recommended under each behavioral target; wherein the behavioral target is used to characterize the interactive behavior expected from the user to the target recommended song.

[0134] The song sorting module 803 is used to sort at least one song to be recommended according to the target characteristics, user characteristics and song characteristics of each song to be recommended under each behavior target, and obtain a sorting result of at least one song to be recommended.

[0135] The song recommendation module 804 is configured to determine a target recommended song corresponding to the song recommendation request from the at least one song to be recommended according to a ranking result of the at least one song to be recommended.

[0136] In one embodiment, the feature extraction module 802 is also used to perform feature fusion processing on the user's user features and the song features of each song to be recommended, so as to obtain a first fusion feature corresponding to each song to be recommended; through the trained multi-objective processing model, the first fusion feature corresponding to each song to be recommended is subjected to feature extraction processing for multiple different behavioral targets, so as to obtain target features of each song to be recommended under each behavioral target; wherein the multi-objective processing model includes at least multiple expert networks, a gating network corresponding to each behavioral target, and a target task tower network corresponding to each behavioral target.

[0137] In one embodiment, the song recommendation device 800 also includes a multi-behavior target feature extraction module, which is used to input the first fusion feature into each expert network in the multiple expert networks respectively to obtain the feature vector output by each expert network; input the first fusion feature into the gating network corresponding to each behavior target respectively to obtain the weight of each expert network under each behavior target; perform weighted summation processing on the weight of each expert network under each behavior target and the feature vector output by each expert network to obtain the fusion feature of the song to be recommended under each behavior target; perform regression processing on the fusion feature of the song to be recommended under each behavior target through the target task tower network corresponding to each behavior target to obtain the target feature of the song to be recommended corresponding to each behavior target.

[0138] In one embodiment, the song recommendation device 800 also includes a user feature acquisition module, which is used to perform feature extraction processing on the user's song sequence based on the attention mechanism to obtain the user's interaction behavior sequence characteristics; and obtain the user's user characteristics based on the user's user attribute characteristics and interaction behavior sequence characteristics.

[0139] In one embodiment, the song sequence includes a user's historically played recommended songs and the user's historically interacted songs. The historically played recommended songs represent at least one song to be recommended that was played by the user within a historical time period. The historically interacted songs represent songs associated with user interactions under various behavioral goals during the historical time period. The song recommendation device 800 also includes a behavior feature acquisition module for performing an attention mechanism on the historically played recommended songs and the historically interacted songs to obtain the user's behavior sequence features, and performing pooling on the behavior sequence features to obtain the user's interaction behavior sequence features.

[0140] In one embodiment, the song sequence includes all songs played in the user's history; the historically played songs represent all songs played by the user within a historical time period, and the historically played songs include the user's historically played recommended songs. The song recommendation device 800 also includes a behavior feature acquisition module for performing feature extraction processing on all songs played in the history to obtain historical song features of all songs played in the history; obtaining an attention weight for each song to be recommended based on the similarity between the song features of each song to be recommended and the historical song features; and performing a weighted summation process on at least one song to be recommended based on the attention weight of each song to be recommended to obtain a user's interactive behavior sequence feature.

[0141] In one embodiment, the song sorting module 803 is further used to perform feature fusion processing on the target features, user features and song features of each song to be recommended under each behavioral target, to obtain a second fused feature of each song to be recommended; and to sort the second fused feature through a pre-trained sorting model to obtain a sorting result of at least one song to be recommended.

[0142] In one embodiment, the song recommendation device 800 also includes a sorting processing module, which is used to perform context information extraction processing on the second fusion feature through the context information processing model in the sorting model to obtain the context information of the second fusion feature; and to perform sorting processing on at least one song to be recommended according to the context information through the multi-objective sorting model in the sorting model to obtain a sorting result of at least one song to be recommended.

[0143] In one embodiment, the song acquisition module 801 is also used to obtain user input data corresponding to the song recommendation request; determine at least one song that matches the user input data from the music library, and obtain at least one song to be recommended corresponding to the song recommendation request.

[0144] Each module in the song recommendation device described above may be implemented in whole or in part through software, hardware, or a combination thereof. Each module may be embedded in or independent of a processor in a computer device in the form of hardware, or may be stored in a memory in the computer device in the form of software, so that the processor can call and execute the corresponding operations of each module.

[0145] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 9As shown. The computer device includes a processor, memory, an input / output interface, a communication interface, a display unit, and an input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and 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 internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals via wired or wireless means, and the wireless means can be achieved via Wi-Fi, mobile cellular networks, NFC (near field communication), or other technologies. When executed by the processor, the computer program implements a song recommendation method. The display unit of the computer device is used to form a visually visible image and can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad set on the computer device casing, or an external keyboard, touchpad or mouse.

[0146] Those skilled in the art will understand that Figure 9 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0147] In one embodiment, a computer device is further provided, including a memory and a processor. The memory stores a computer program, and the processor implements the steps in the above method embodiments when executing the computer program.

[0148] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0149] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the above method embodiments when executed by a processor.

[0150] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. 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 above-mentioned embodiments. In particular, any reference to memory, database, 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 various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The databases involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchains. The processors involved in the various embodiments provided herein may be, but are not limited to, general-purpose processors, central processing units (CPUs), graphics processing units (GPUs), digital signal processors (DSPs), programmable logic devices (PLDs), data processing logic devices based on quantum computing, and the like.

[0151] The technical features of the above embodiments can 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.

[0152] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. A song recommendation method, characterized in that: The method comprises: In response to a song recommendation request from a user, obtaining at least one song to be recommended corresponding to the song recommendation request; Performing feature extraction processing for multiple different behavioral targets on the user features of the user and the song features of each of the songs to be recommended, to obtain target features of each of the songs to be recommended under each of the behavioral targets; wherein the behavioral target is used to characterize the interactive behavior that the user is expected to perform on the target recommended song; Sorting the at least one song to be recommended according to the target feature of each song to be recommended under each behavior target, the user feature, and the song feature of each song to be recommended, to obtain a ranking result of the at least one song to be recommended; According to the ranking result of the at least one song to be recommended, a target recommended song corresponding to the song recommendation request is determined from the at least one song to be recommended.

2. The method according to claim 1, characterized in that The step of performing feature extraction processing on the user features of the user and the song features of each song to be recommended for a plurality of different behavioral targets to obtain target features of each song to be recommended under each behavioral target includes: Performing feature fusion processing on the user feature of the user and the song feature of each of the songs to be recommended to obtain a first fusion feature corresponding to each of the songs to be recommended; Through the trained multi-objective processing model, the first fusion feature corresponding to each of the songs to be recommended is subjected to feature extraction processing for multiple different behavioral targets to obtain the target features of each of the songs to be recommended under each of the behavioral targets; wherein, the multi-objective processing model includes at least multiple expert networks, a gating network corresponding to each of the behavioral targets, and a target task tower network corresponding to each of the behavioral targets.

3. The method according to claim 2, characterized in that The multi-objective processing model completed through training performs feature extraction processing of multiple different behavioral objectives on the first fusion feature corresponding to each of the songs to be recommended, and obtains the target features of each of the songs to be recommended under each of the behavioral objectives, including: Inputting the first fusion feature into each of the multiple expert networks respectively to obtain a feature vector output by each of the expert networks; Inputting the first fusion features into the gating network corresponding to each of the behavioral targets respectively to obtain the weight of each of the expert networks under each of the behavioral targets; Performing weighted summation processing on the weight of each expert network under each behavioral target and the feature vector output by each expert network to obtain the fusion features of the song to be recommended under each behavioral target; Through the target task tower network corresponding to each behavioral target, the fusion features of the song to be recommended under each behavioral target are regressed to obtain the target features of the song to be recommended corresponding to each behavioral target.

4. The method according to claim 1, wherein Before performing feature extraction processing for a plurality of different behavioral targets on the user features and the song features of each song to be recommended to obtain the target features of each song to be recommended under each behavioral target, the method further includes: Performing feature extraction processing on the user's song sequence based on the attention mechanism to obtain the user's interactive behavior sequence features; The user characteristics of the user are obtained according to the user attribute characteristics of the user and the interaction behavior sequence characteristics.

5. The method according to claim 4, characterized in that The song sequence includes the user's historical play recommendation songs and the user's historical interaction songs; wherein the historical play recommendation songs represent songs in the at least one song to be recommended that have been played by the user within a historical time period; The historical interactive songs represent songs associated with the interactive behaviors under the multiple different behavioral goals implemented by the user within a historical time period; The feature extraction process of the user's song sequence based on the attention mechanism to obtain the user's interactive behavior sequence features includes: Performing attention mechanism processing on the historical play recommendation songs and the historical interactive songs to obtain the user's behavior sequence features; Pooling is performed on the behavior sequence features to obtain the user's interactive behavior sequence features.

6. The method according to claim 4, characterized in that The song sequence includes all songs played in the history of the user; the songs played in the history represent all songs played by the user in a historical time period, and the songs played in the history include the user's recommended songs; The feature extraction process of the user's song sequence based on the attention mechanism to obtain the user's interactive behavior sequence features includes: Performing feature extraction processing on all the songs played in the history to obtain historical song features of all the songs played in the history; Obtaining an attention weight for each song to be recommended based on the similarity between the song features of each song to be recommended and the features of the historically played songs; A weighted summation process is performed on at least one of the songs to be recommended according to the attention weight of each of the songs to be recommended to obtain the interactive behavior sequence feature of the user.

7. The method according to claim 1, characterized in that Sorting the at least one song to be recommended according to the target feature of each song to be recommended under each behavior target, the user feature, and the song feature of each song to be recommended, to obtain a ranking result of the at least one song to be recommended, includes: Performing feature fusion processing on the target feature of each song to be recommended under each behavior target, the user feature, and the song feature of each song to be recommended to obtain a second fusion feature of each song to be recommended; The second fusion feature is sorted using a pre-trained sorting model to obtain a sorting result of the at least one song to be recommended.

8. The method according to claim 7, characterized in that The step of sorting the second fusion features using a pre-trained sorting model to obtain a sorting result of the at least one song to be recommended includes: Performing context information extraction processing on the second fused feature through a context information processing model in the sorting model to obtain context information of the second fused feature; The at least one song to be recommended is sorted according to the context information through the multi-objective sorting model in the sorting model to obtain a sorting result of the at least one song to be recommended.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

11. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images