Multi-scene switching method, system, device and medium based on a song request station

Abstract

The application belongs to the field of human-computer interaction, and specifically discloses a multi-scene switching method, system, device and medium based on a song request station, wherein the method comprises collecting, through a sensor, user stay time, song request frequency and interaction intensity behavior data in a virtual environment, processing the behavior data by using a long short-term memory network to obtain a user dynamic preference vector; matching the user dynamic preference vector with a preset scene type library, and if the matching degree is higher than a threshold, judging that the current scene preference is stable, otherwise judging that a potential switching intention exists, and obtaining a potential switching intention probability distribution through the above judgment; after obtaining the potential switching intention probability distribution, training the correlation characteristics between multi-scene data by using a support vector machine classifier; and the application aims to solve the problem in the prior art that the rapid change of user scene preference in a virtual environment leads to resource loading delay and unsmooth interface switching.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of human-computer interaction, and particularly relates to a multi-scene switching method, system, device and medium based on a song request station. BACKGROUND

[0002] As a popular online entertainment form, the song request station attracts a large number of users to interact with music through virtual scenes. Its core is to provide users with immersive and personalized scene experience. Research on how to optimize the scene switching of the song request station through user behavior prediction technology not only concerns the smoothness of user experience, but also directly affects the resource allocation efficiency and operation stability of the platform. With the growth of the user scale and the increase of scene diversity, accurately predicting the behavior tendency of users between different scenes has become a key field to improve the service quality of the platform and the satisfaction of users. The importance of this field lies in that it can capture the dynamic needs of users through technical means, and then realize intelligent allocation of resources and seamless connection of interfaces, so as to bring users a more natural interactive experience. However, the existing methods often face the problem of being difficult to capture the dynamic preferences of users when dealing with user behavior prediction. Many schemes rely on simple user portraits or static historical data analysis, ignoring the variability and real-time nature of user behavior in the song request station scene. For example, users may temporarily change their preferences due to their current mood, social interaction or scene atmosphere, but existing methods are difficult to quickly adapt to such dynamic changes. In addition, although some schemes attempt to use real-time data, they are often limited to the analysis of a single scene, lacking a comprehensive understanding of the user's switching behavior between different scenes. This leads to the system being unable to accurately determine the scene that the user may enter next, thereby affecting the efficiency of resource allocation and the coherence of user experience. In the song request station scene, the core technical difficulty first lies in how to accurately quantify the scene preferences of users. SUMMARY

[0003] The application provides a multi-scene switching method, system, device and medium based on a song request station, which aims to solve the problem of resource loading delay and unsmooth interface switching caused by the rapid change of user scene preferences in a virtual environment in the prior art.

[0004] To solve the above technical problems, the technical scheme adopted by the application is:

[0005] The multi-scene switching method based on a song request platform comprises: collecting, through a sensor, user staying time, song request frequency and interaction intensity behavior data in a virtual environment, processing the behavior data by using a long short-term memory network to obtain a user dynamic preference vector; matching the user dynamic preference vector with a preset scene type library, if the matching degree is higher than a threshold, it is judged that the current scene preference is stable, otherwise it is judged that a potential switching intention exists, and the potential switching intention probability distribution is obtained through the above judgment; after obtaining the potential switching intention probability distribution, the correlation features between multi-scene data are trained by using a support vector machine classifier to obtain a possible target scene list of cross-scene switching; the target scene with the highest probability in the possible target scene list is extracted, and a real-time monitored behavior data change trend is combined to judge, if the change trend points to a chorus area, it is determined that the chorus area will be switched to, otherwise the monitoring is continued to obtain an exact switching demand prediction result; according to the exact switching demand prediction result, the resource files and interface elements of the target scene are preloaded, a dynamic allocation algorithm is used to adjust the server bandwidth allocation ratio to obtain an optimized resource configuration scheme; the interface layout parameters are updated through the optimized resource configuration scheme, if the prediction result matches the actual user operation, the layout parameters are directly applied, otherwise the original configuration is rolled back to determine the final interface seamless connection state; feedback data is extracted in the final interface seamless connection state and input into the iterative training process of the long short-term memory network to update the model parameters to adapt to the user behavior mode.

[0006] In an aspect of the present disclosure, the step of collecting, through a sensor, user staying time, song request frequency and interaction intensity behavior data in a virtual environment, processing the behavior data by using a long short-term memory network to obtain a user dynamic preference vector comprises:

[0007] User behavior data in a virtual environment is collected by a sensor, including staying time, song request frequency and interaction intensity, and stored as structured time series data to obtain an initial behavior data set;

[0008] Data cleaning techniques are used to pre-process the initial behavior data set to remove noise data and missing values, and the staying time, song request frequency and interaction intensity are aligned by timestamp to obtain standardized behavior sequences;

[0009] The standardized behavior sequences are time series modeled according to a long short-term memory network to extract time-dependent features and generate a user dynamic preference vector;

[0010] According to the user dynamic preference vector, a clustering algorithm is used to group users to obtain user preference categories based on the similarity of behavior mode vectors;

[0011] If the user preference category matches the preset music recommendation rule, a corresponding song list is obtained from the music library according to the category characteristics to generate a personalized recommendation sequence;

[0012] According to the comparison between the personalized recommendation sequence and the user song ordering frequency, the matching degree of the recommendation sequence and the actual behavior is calculated to obtain a recommendation effect evaluation index;

[0013] According to the recommendation effect evaluation index, the long short-term memory network parameters are optimized by using a gradient descent algorithm to update the user dynamic preference vector.

[0014] In an aspect of the present disclosure, the step of matching the user dynamic preference vector with a preset scene type library, and if the matching degree is higher than a threshold, judging that the current scene preference is stable, otherwise judging that the potential switching intention exists, and obtaining a potential switching intention probability distribution through the above judgment, comprises:

[0015] Obtain the user dynamic preference vector, generate a feature vector by analyzing the user historical interaction data;

[0016] Based on the vector matching algorithm, compare the user dynamic preference vector with the standard vector in the preset scene type library to obtain a matching score;

[0017] If the matching score is higher than a preset threshold, it is judged that the current scene preference is stable, and a stable scene label is generated;

[0018] According to the stable scene label, extract the recommendation content related to the current scene to generate a personalized recommendation list;

[0019] If the matching score is lower than a preset threshold, analyze the potential switching intention by using a Bayesian inference algorithm to obtain an intention probability distribution;

[0020] According to the intention probability distribution, a weighted ranking algorithm is used to generate a recommendation sequence of the potential switching scene;

[0021] According to the recommendation sequence, dynamically adjust the user interaction interface, and output the optimized scene recommendation content.

[0022] In an aspect of the present disclosure, after obtaining the potential switching intention probability distribution, the step of using a support vector machine classifier to train the correlation characteristics between multi-scene data to obtain a possible target scene list of cross-scene switching comprises:

[0023] Obtain the user interaction data, and extract the probability distribution of the potential switching intention from the user interaction data;

[0024] Through probability distribution analysis, a support vector machine classifier is used to train the correlation characteristics of multi-scene data to obtain a classification model;

[0025] If the prediction accuracy of the classification model reaches the preset threshold, a target scene list for cross-scene switching is determined according to the model output;

[0026] According to the target scene list, an associated feature set of each scene is extracted to generate a feature vector;

[0027] The feature vector is used to group the target scenes by a clustering algorithm to obtain a scene grouping result;

[0028] The priority of each group is obtained through the scene grouping result to obtain a final switching scene sequence;

[0029] If the priority of the final switching scene sequence meets the preset condition, the scene sequence is output, otherwise the feature vector is re-extracted for clustering.

[0030] In an aspect of the present disclosure, the step of extracting the target scene with the highest probability from the possible target scene list and combining the real-time monitored behavior data change trend to determine whether the change trend points to the chorus area to determine that the chorus area will be switched to, otherwise continue to monitor to obtain the exact switching demand prediction result, comprises:

[0031] The target scene is obtained from the scene list, the highest probability target scene is extracted by using a probability sorting algorithm to obtain a highest probability target scene identifier;

[0032] The behavior data is collected by a real-time monitoring system, and a time series analysis method is used to calculate the change trend of the behavior data to obtain a behavior data change trend vector;

[0033] If the behavior data change trend vector points to the preset feature of the chorus area, the target scene is determined to be the chorus area, and a scene switching signal is generated;

[0034] According to the scene switching signal, the configuration parameters related to the chorus area are obtained to determine the switching demand;

[0035] If the behavior data change trend vector does not point to the preset feature of the chorus area, the behavior data is updated by a continuous monitoring module to obtain a new behavior data change trend vector;

[0036] The new behavior data change trend vector is classified by using a random forest algorithm to determine whether the chorus area switching condition is met to obtain a classification result;

[0037] According to the classification result, a final switching demand prediction result is determined to generate a target scene switching instruction.

[0038] In an aspect of the present disclosure, the step of preloading resource files and interface elements of the target scene according to the prediction result of the exact switching requirement, adjusting the server bandwidth allocation ratio by a dynamic allocation algorithm, and obtaining an optimized resource configuration scheme comprises:

[0039] The prediction result of the switching requirement is obtained by analyzing user behavior data through a prediction model;

[0040] Based on the prediction result, the priority and resource requirement of the target scene are determined, and a resource loading sequence is generated;

[0041] According to the resource loading sequence, the resource files and interface elements of the target scene are preloaded and stored in the local cache;

[0042] A consistent hashing algorithm is used to dynamically allocate server bandwidth and adjust the bandwidth ratio of each scene;

[0043] If the loading speed of a scene is lower than a preset threshold after adjusting the bandwidth ratio, the bandwidth is redistributed through a load balancing mechanism to obtain an optimized bandwidth allocation scheme;

[0044] According to the optimized bandwidth allocation scheme, the resource loading state is monitored in real time to obtain loading completion data;

[0045] According to the loading completion data, the triggering time of the preloading mechanism is dynamically adjusted to generate a final resource configuration scheme.

[0046] In an aspect of the present disclosure, the step of updating the interface layout parameters through the optimized resource configuration scheme, directly applying the layout parameters if the prediction result matches the actual user operation, otherwise rolling back to the original configuration, and determining the final seamless interface connection state comprises:

[0047] The real-time allocation state of the resource configuration scheme is determined by analyzing the resource configuration scheme to obtain the currently available computing resources and storage resources from a preset resource pool;

[0048] According to the real-time allocation state of the resource configuration scheme, a preset interface rendering algorithm is used to generate corresponding interface layout parameters to obtain preliminarily adjusted layout parameters;

[0049] The preliminarily adjusted layout parameters are predicted through a machine learning classification model to obtain a prediction result, and it is judged whether the prediction result conforms to the expected user operation mode;

[0050] If the prediction result is consistent with the actual user operation, the preliminarily adjusted layout parameters are deployed to the interface rendering module through parameter application logic to determine the update state of the interface layout;

[0051] If the prediction result is inconsistent with the actual user operation, the original layout parameters are obtained from the original layout parameter repository, a rollback operation mechanism is adopted to restore the interface to the initial state, and a rolled-back interface state is obtained;

[0052] According to the rolled-back interface state or the updated state, a state synchronization algorithm is adopted to detect the delay and consistency in the interface switching process, and a seamless switching state is determined;

[0053] By recording the seamless switching state and user operation data, a training data set of a dynamic adjustment mechanism is updated, an optimized resource configuration scheme is obtained, and input parameters for the next adjustment are obtained.

[0054] In another aspect of the present disclosure, the present disclosure also relates to a multi-scene switching system based on a song ordering platform, characterized in that it comprises:

[0055] A behavior data processing module is configured to collect user stay time, song ordering frequency and interaction intensity behavior data in a virtual environment through a sensor, process the behavior data using a long short-term memory network, and obtain a user dynamic preference vector;

[0056] A scene preference judgment module is configured to match the user dynamic preference vector with a preset scene type library, and if the matching degree is higher than a threshold, it is judged that the current scene preference is stable, otherwise it is judged that there is a potential switching intention, and the potential switching intention probability distribution is obtained through the above judgment;

[0057] An associated feature training module is configured to obtain the potential switching intention probability distribution, train the associated features between multi-scene data using a support vector machine classifier, and obtain a possible target scene list for cross-scene switching;

[0058] A switching demand prediction module is configured to extract the target scene with the highest probability from the possible target scene list, and combine the real-time monitored behavior data change trend to determine whether the change trend points to a chorus area, and if so, determine that the chorus area will be switched to, otherwise continue to monitor, and obtain an exact switching demand prediction result;

[0059] A resource configuration optimization module is configured to preload resource files and interface elements of the target scene according to the exact switching demand prediction result, adjust the server bandwidth allocation ratio using a dynamic allocation algorithm, and obtain an optimized resource configuration scheme;

[0060] An interface connection determination module is configured to update the interface layout parameters through the optimized resource configuration scheme, and if the prediction result matches the actual user operation, the layout parameters are directly applied, otherwise the original configuration is rolled back, and the final interface seamless connection state is determined;

[0061] The model iterative updating module is configured to extract feedback data in the final interface seamless connection state, input the feedback data into an iterative training process of the long short-term memory network, and update model parameters to adapt to user behavior patterns.

[0062] In another aspect of the present disclosure, the present disclosure also relates to a jukebox device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the jukebox-based multi-scene switching method described above when executing the computer program.

[0063] In another aspect of the present disclosure, the present disclosure also relates to a computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the steps of the jukebox-based multi-scene switching method described above when executed by a processor.

[0064] Compared with the prior art, the present application has the following beneficial effects:

[0065] The present application realizes seamless scene switching by fusing behavior data analysis, dynamic preference modeling, and resource optimization configuration. First, the long short-term memory network is used to process user dwell time, song ordering frequency, and interaction intensity behavior data to generate a dynamic preference vector, which is matched with a scene type library to determine preference stability and potential switching intention. Then, the support vector machine classifier is used to train cross-scene correlation features to determine a possible target scene list, and the specific switching demand is predicted in combination with real-time behavior data trends. For the prediction result, the target scene resources are preloaded, the dynamic allocation algorithm is used to optimize the server bandwidth, the interface layout parameters are finally updated, and the actual operation is verified or rolled back. The present application iteratively updates the model parameters through feedback data, significantly improves the fluency and resource utilization efficiency of scene switching, and realizes personalized and real-time response virtual environment experience. BRIEF DESCRIPTION OF DRAWINGS

[0066] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be considered as a limitation to the scope. For those skilled in the art, other related drawings can also be obtained without creative labor.

[0067] Figure 1 The flowchart of the jukebox multi-scene switching method in the present application.

[0068] Figure 2 The schematic diagram of the jukebox multi-scene switching method in the present application. DETAILED DESCRIPTION

[0069] The application will be described further below with reference to examples. The examples described are only a part of the examples of the application and are not all the examples. Based on the examples in the application, all other examples obtained by those of ordinary skill in the art without creative work are within the protection scope of the application.

[0070] Example one

[0071] Please refer to Figure 1 and Figure 2 It is disclosed that the multi-scene switching method based on the karaoke machine includes: collecting the user's stay time, song frequency and interactive intensity behavior data in the virtual environment through the sensor, using the long short-term memory network to process the behavior data, and obtaining the user dynamic preference vector; matching the user dynamic preference vector with the preset scene type library, if the matching degree is higher than the threshold, it is judged that the current scene preference is stable, otherwise it is judged that the potential switching intention exists, and the potential switching intention probability distribution is obtained through the above judgment; after obtaining the potential switching intention probability distribution, the support vector machine classifier is used to train the correlation characteristics between the multi-scene data, and the possible target scene list of cross-scene switching is obtained; the highest probability target scene is extracted in the possible target scene list, and the real-time monitored behavior data change trend is combined to judge, if the change trend points to the chorus area, it is determined that it will be switched to the chorus area, otherwise it continues to be monitored, and the exact switching demand prediction result is obtained; according to the exact switching demand prediction result, the resource files and interface elements of the target scene are preloaded, the dynamic allocation algorithm is used to adjust the server bandwidth allocation ratio, and the optimized resource configuration scheme is obtained; the interface layout parameters are updated through the optimized resource configuration scheme, if the prediction result matches the actual user operation, the layout parameters are directly applied, otherwise it is rolled back to the original configuration, and the final interface seamless connection state is determined; the feedback data is extracted in the final interface seamless connection state, input to the iterative training process of the long short-term memory network, the model parameters are updated to adapt to the user behavior mode.

[0072] This invention achieves seamless scene switching by integrating behavioral data analysis, dynamic preference modeling, and resource optimization. First, it utilizes a Long Short-Term Memory (LSTM) network to process user dwell time, song selection frequency, and interaction intensity data, generating a dynamic preference vector. This vector is then matched against a scene type library to determine preference stability and potential switching intentions. Next, a Support Vector Machine (SVM) classifier is used to train cross-scene association features, identifying a list of possible target scenes. Specific switching needs are predicted based on real-time behavioral data trends. For the prediction results, this invention preloads target scene resources, employs a dynamic allocation algorithm to optimize server bandwidth, and finally updates interface layout parameters, verifying or rolling back the configuration based on actual operation. By iteratively updating model parameters through feedback data, this invention significantly improves the smoothness of scene switching and resource utilization efficiency, achieving a personalized, real-time responsive virtual environment experience.

[0073] Example 2

[0074] Please see Figure 1 and Figure 2 As shown, this embodiment is a further optimization based on embodiment one. In this embodiment, a multi-scene switching method based on a karaoke system is disclosed, which may specifically include:

[0075] This invention provides a personalized music recommendation and scene switching method based on user behavior data. It aims to generate dynamic preferences and achieve accurate prediction and optimized resource allocation for scene switching by analyzing user behavior data in a virtual environment. The method is described in detail below with reference to specific implementation methods.

[0076] Step S1: Collect user behavior data such as dwell time, song selection frequency, and interaction intensity in the virtual environment through sensors to generate an initial behavior dataset.

[0077] Specifically, a virtual environment can be an interactive scenario based on virtual reality (VR) or augmented reality (AR) devices. For example, a user enters a virtual karaoke room through a head-mounted display (HUD) to select songs and interact with other users. Sensors, including but not limited to accelerometers, gyroscopes, eye trackers, or microphone arrays in the HUD, are used to capture real-time behavioral data of the user in the virtual environment. Dwell time refers to the duration a user remains in a specific virtual scene, such as a song selection interface or a virtual stage, typically recorded in seconds. Song selection frequency refers to the number of times a user selects a song per unit of time, such as the number of song selections per minute. Interaction intensity is measured by the frequency of the user's gestures, voice commands, or the use of virtual props, such as the number of interactions between the user and virtual characters in a virtual karaoke room or changes in the volume of voice commands.

[0078] In one possible implementation, the sensor collects data at a fixed frequency, such as 10 times per second, and stores the collected data as a structured time series with timestamps. For example, a user stays in a virtual KTV room for 5 minutes, orders songs 3 times, and interacts with the virtual scene 10 times through gestures during this period. The sensor records these data in chronological order as an initial behavior data set. It should be noted that the initial behavior data set may contain noise, such as abnormal gesture data caused by device jitter or timestamp deviation caused by network delay. To ensure data quality, the sensor data can be preliminarily processed through filtering techniques during the collection process to remove obvious outliers and generate a more reliable initial behavior data set.

[0079] In step S11, the initial behavior data set is preprocessed to remove noise data and missing values to obtain a standardized behavior sequence.

[0080] Specifically, the initial behavior data set may contain noise data caused by device failure or user misoperation, such as incomplete song ordering record due to network interruption or empty stay time record for a certain period. The preprocessing process first identifies the empty values or incomplete records in the data through a missing value detection algorithm, such as checking whether the timestamps are continuous. If a certain period of timestamps is missing, the data is completed by interpolation method, such as filling the missing stay time based on the average value of the previous and subsequent timestamps.

[0081] The removal of noise data can be achieved by statistical methods, such as calculating the mean and standard deviation of the song ordering frequency within a time window. If a certain song ordering frequency exceeds the range of mean plus or minus three times the standard deviation, it is considered as noise and removed. In addition, to ensure the comparability of different types of data, the stay time, song ordering frequency and interaction intensity need to be normalized to map them to the same numerical interval, such as 0 to 1.

[0082] For example, the stay time can be converted to a relative duration ratio through linear normalization, the song ordering frequency can be normalized according to the maximum frequency, and the interaction intensity can be standardized based on the intensity value of gestures or speech. After preprocessing, the behavior sequence with timestamp alignment and numerical standardization is obtained, which contains stay time, song ordering frequency and interaction intensity data arranged in chronological order.

[0083] In step S12, the standardized behavior sequence is time series modeled by a long short-term memory network to extract time-dependent features and generate a user dynamic preference vector.

[0084] The long short-term memory network is a neural network suitable for processing time series data and can capture long-term dependencies in behavior sequences. For example, a user may order songs with low frequency first and then high frequency in a virtual KTV due to emotional changes, and this time series pattern can be modeled by a long short-term memory network.

[0085] In an embodiment, the standardized behavior sequence is input into a long short-term memory network at time steps, each time step containing a three-dimensional vector of dwell time, song ordering frequency, and interaction intensity. The network determines which historical information to retain and updates the current state through forget gate, input gate, and output gate mechanisms. For example, the forget gate can filter out irrelevant data at earlier time steps, such as random operations when the user initially enters the virtual environment, the input gate focuses on retaining recent high-frequency song ordering behavior, and the output gate generates a behavior feature vector at the current time step.

[0086] It should be noted that the number of hidden layer nodes of the long short-term memory network can be adjusted according to the complexity of the virtual environment. For example, in a simple virtual KTV scenario, fewer hidden nodes, such as 64, can be set to improve computational efficiency; in a complex multi-scene virtual environment, the number of nodes can be increased, such as 256, to capture more rich temporal features. After processing by multiple layers of the network, a fixed-dimensional user dynamic preference vector, such as a 128-dimensional vector, is generated to represent the user's behavior pattern in the virtual environment.

[0087] For example, a user continuously selects fast-paced songs in a virtual KTV with high interaction intensity, and the network can capture the pattern of his preference for fast-paced music, generating a corresponding behavior pattern vector. This vector not only reflects the user's current behavior, but also predicts his potential preference trend through temporal modeling.

[0088] Step S2, according to the user dynamic preference vector, a clustering algorithm is used to group users and determine user preference categories.

[0089] Specifically, the clustering algorithm is used to classify users with similar behavior patterns into the same category to provide personalized recommendations for different preference users. The clustering algorithm can be a distance-based K-means algorithm or a density-based DBSCAN algorithm.

[0090] In a possible implementation, the K-means algorithm is used to group the user dynamic preference vectors. First, the number of clusters K is determined, for example, according to the user behavior characteristics of the virtual KTV, K is preset to 4, representing four categories of users: fast-paced preference, romantic preference, random song ordering, and low interaction.

[0091] The algorithm assigns users to the nearest cluster center by calculating the Euclidean distance between the behavior pattern vectors.

[0092] For example, a user dynamic preference vector shows that he has a high song ordering frequency and prefers fast-paced songs, and the algorithm classifies him into the fast-paced preference category; while another user has a long dwell time but a low song ordering frequency and weak interaction intensity, he may be classified into the low interaction category. It should be noted that the initialization of the cluster center can be optimized through random selection or prior knowledge based on historical data to improve the accuracy of grouping.

[0093] In another embodiment, if the user behavior in the virtual KTV scene is complex, the DBSCAN algorithm can be used to automatically identify the density distribution of the behavior mode vector and generate a dynamic number of clustering categories; this method is suitable for processing irregularly distributed user behavior data, such as some users who sing songs in a specific time period.

[0094] Finally, the clustering result generates a user preference category label, such as fast rhythm preference or emotional preference, which provides a basis for subsequent recommendation.

[0095] Step S21, if the user preference category matches the preset music recommendation rule, the corresponding song list is obtained from the music library, and a personalized recommendation sequence is generated.

[0096] Specifically, the music recommendation rule is a pre-defined mapping relationship between the category and the song features. For example, the fast rhythm preference category corresponds to songs with a rhythm of more than 120 beats per minute, and the emotional preference category corresponds to songs with a slow rhythm and rich lyrics. Each song in the music library is pre-labeled with feature tags such as rhythm, style, and emotional attributes.

[0097] In one embodiment, if a user is classified as a fast rhythm preference category, the system filters songs with a rhythm of more than 120 beats per minute from the music library according to the music recommendation rule, and generates a recommendation sequence containing 10 songs. The recommendation sequence is sorted by the popularity of the song or the user's historical song record to improve the acceptance of the recommendation.

[0098] For example, the system can preferentially recommend songs of the same singer or style as the user recently sang.

[0099] It should be noted that the music recommendation rule can be dynamically adjusted according to the virtual KTV scene. For example, in a multi-person chorus scene, the rule can preferentially recommend popular songs suitable for chorus, while in a single-person song selection scene, more attention is paid to personal preferences. The generation of personalized recommendation sequences improves the user's immersion and satisfaction in the virtual environment.

[0100] Step S22, by comparing the personalized recommendation sequence with the user's song selection frequency, the matching degree of the recommendation sequence and the actual behavior is calculated, and the recommendation effect evaluation index is obtained.

[0101] Specifically, the matching degree is calculated by comparing the similarity of the songs in the recommendation sequence and the user's actual song selection record. For example, the system counts the features of the songs in the recommendation sequence, such as rhythm, style, and the overlap ratio of the features of the user's recent song selection record.

[0102] In one possible implementation, the matching degree is calculated by cosine similarity. The feature vectors of the recommended sequence and the user's song selection records are compared to generate a similarity score. For example, if the recommended sequence contains 5 fast-paced songs and the user's recent song selection records contain 4 fast-paced songs, the matching degree is high. The matching degree score is further converted into a recommendation effect evaluation index, such as a standardized score from 0 to 1, and the higher the score, the more the recommendation conforms to the user's preferences.

[0103] In another embodiment, if the virtual KTV scene supports multi-person interaction, a weighted matching degree calculation can be introduced to consider both the user's personal song selection records and the interaction behavior with the virtual role.

[0104] For example, the user prefers a certain style when singing with the virtual role, and the system increases the weight of that style to generate more accurate evaluation indexes. This method effectively improves the adaptability of the recommendation system in complex scenarios.

[0105] Step S23, according to the recommendation effect evaluation index, the gradient descent algorithm is used to optimize the long short-term memory network parameters, and the user dynamic preference vector is updated.

[0106] Specifically, the recommendation effect evaluation index reflects the accuracy of the recommended sequence. If the index is low, it means that the behavior pattern vector generated by the long short-term memory network fails to accurately capture the user's preferences. The gradient descent algorithm optimizes the network parameters by minimizing the loss function, such as mean square error.

[0107] In one embodiment, the system uses the recommendation effect evaluation index as the input of the loss function to calculate the deviation between the network's predicted user dynamic preference vector and the actual behavior. The gradient descent algorithm adjusts the weights and biases of the network, such as updating the parameters of the forget gate and the input gate, so that the network is more inclined to retain the behavior characteristics related to high matching degree recommendations.

[0108] For example, a user recently prefers ballads, but the network generates a recommended sequence that is biased towards fast-paced songs, resulting in a low matching degree. Through multiple iterations, the gradient descent algorithm adjusts the network parameters to more accurately capture the user's preference for ballads and generates an updated behavior pattern vector.

[0109] In another possible implementation, a learning rate adaptive adjustment mechanism can be introduced, such as dynamically reducing the learning rate when the loss function decreases slowly to improve the optimization accuracy. This method is particularly effective in the dynamic environment of virtual KTV, and can quickly adapt to changes in user behavior.

[0110] Step S3, according to the user dynamic preference vector and the preset scene type library, it is judged whether the current scene preference is stable or whether there is a potential switching intention, and a potential switching intention probability distribution is obtained.

[0111] Specifically, the user dynamic preference vector is generated by a long short-term memory network, representing the user's behavioral tendency in the current virtual environment. The scene type library contains standard vectors of various virtual KTV scenes, such as single song ordering, multi-person singing, and virtual stage performance, etc. Each scene has pre-defined typical behavior characteristics, such as high song ordering frequency and low interaction intensity.

[0112] In one embodiment, the system calculates the similarity between the user dynamic preference vector and each standard vector in the scene type library through a vector matching algorithm. For example, using cosine similarity calculation, if the similarity between the user dynamic preference vector and the standard vector of the multi-person singing scene is higher than the preset threshold 0.8, it is determined that the current scene preference is stable, and a multi-person singing scene label is generated. If the similarity is lower than the threshold, it means that the user may have the intention to switch to other scenes.

[0113] In another possible implementation, the system analyzes the potential switching intention through a Bayesian inference algorithm. For example, based on the user's historical interaction data, the conditional probability of the user switching from the current scene to other scenes is calculated, generating an intention probability distribution. For example, the probability distribution may show that the user has a 60% probability of switching to a virtual stage performance scene and a 30% probability of remaining in a multi-person singing scene.

[0114] It should be noted that the scene type library can be dynamically expanded according to the business needs of virtual KTV, for example, a virtual DJ mixing scene is added, and the system will update the standard vector and recalculate the matching degree. This flexibility ensures that the recommendation system can adapt to the needs of different user groups.

[0115] So far, the first half of the method realizes personalized recommendation and preliminary scene preference judgment by collecting user behavior data, generating a dynamic preference vector, and performing scene matching, laying a foundation for subsequent scene switching and resource optimization.

[0116] Step S4, after obtaining the potential switching intention probability distribution, a support vector machine classifier is used to train the correlation features between multiple scene data to determine the possible target scene list for cross-scene switching.

[0117] Specifically, the potential switching intention probability distribution is generated by step S3, representing the possibility of the user switching from the current virtual KTV scene to other scenes. The support vector machine classifier analyzes the correlation features between multiple scene data to build a classification model for predicting possible switching target scenes.

[0118] For example, in a virtual KTV, when a user switches from a single song ordering scene to a multi-person singing scene, the behavior characteristics involved may include increased song ordering frequency or increased interaction intensity.

[0119] In one possible implementation, the system extracts probability distributions of potential switching intentions from user interaction data, such as a probability distribution vector of three scenarios including solo karaoke, group singing, and virtual stage performance. A support vector machine classifier takes these probability distributions as input features, combines with behavioral features in historical interaction data such as dwell time, karaoke frequency, and interaction intensity, and trains a classification model; during the training process, the classifier generates a decision boundary that can distinguish different scenarios by maximizing the separation between features in different scenarios. For example, if a user has a short dwell time but an increased interaction intensity in the solo karaoke scenario, the classifier may predict that the user tends to switch to the group singing scenario.

[0120] It should be noted that the training data includes user behavior samples in multiple scenarios, such as thousands of user interaction records extracted from the logs of a virtual KTV platform. The sample data is labeled by scenario to ensure that the classifier can learn the feature distribution of different scenarios. After training, the classifier outputs a list of possible target scenarios, such as a list containing two high-probability scenarios of group singing and virtual stage performance. This method improves the accuracy of cross-scenario switching prediction and provides a reliable basis for subsequent scenario grouping and resource preloading.

[0121] Step S41, the associated features of multi-scenario data are trained by a support vector machine classifier through probability distribution analysis, and a classification model is obtained.

[0122] Specifically, the probability distribution analysis is based on the intention probability distribution generated in step S3, and combines the associated features of multi-scenario data, such as the karaoke preference, interaction frequency, or dwell time change pattern of users in different scenarios. The support vector machine classifier maps the non-linear features to a high-dimensional space through a kernel function, such as a radial basis function, to generate a classification model.

[0123] In one embodiment, the system extracts feature data from the interaction logs of a virtual KTV platform, such as a user's karaoke frequency of 1 time per minute and low interaction intensity in the solo karaoke scenario, and an increased karaoke frequency of 3 times per minute and significantly improved interaction intensity in the group singing scenario. The classifier learns the differences between scenarios by analyzing these features and generates a decision boundary that can distinguish between solo karaoke and group singing. During the training process, the system can adjust the hyperparameters of the classifier, such as the regularization parameter C, through cross-validation to improve the generalization ability of the model.

[0124] In another possible implementation, if there are many types of virtual KTV scenarios, a multi-class support vector machine can be introduced to train each scenario combination separately. For example, for the switching from "solo karaoke" to "virtual stage performance", the classifier focuses on analyzing the increased frequency of virtual props used by users in the performance scenario. Finally, the classification model outputs the prediction probability of each scenario to form the basis of the target scenario list.

[0125] Step S42, if the prediction accuracy of the classification model reaches the preset threshold, determine the target scene list for cross-scene switching according to the model output.

[0126] Specifically, the prediction accuracy of the classification model is evaluated by the validation set, for example, the classification accuracy of the model is calculated on the test data. If the accuracy reaches the preset threshold, such as 0.85, the system generates a target scene list for cross-scene switching according to the probability distribution output by the model. For example, the list may contain three scenes: multi-person chorus, virtual stage performance, and virtual DJ mixing, sorted in descending order of predicted probability.

[0127] In one embodiment, the system evaluates the classification model in the test environment of the virtual KTV platform, using 1000 user interaction records as the validation set. If the model correctly predicts more than 85% of the scene switching instances, the model is considered reliable and outputs the target scene list.

[0128] For example, a user is currently in a solo karaoke scene, the model predicts that he has a 70% probability of switching to a multi-person chorus and a 20% probability of switching to a virtual stage performance, and the system generates a list containing these two scenes.

[0129] It should be noted that if the accuracy does not reach the threshold, the system can retrain by increasing the training samples or adjusting the classifier parameters, for example, supplementing more interaction data of virtual DJ mixing scenes to improve the model's prediction ability for new scenes. This method ensures the accuracy of the target scene list and provides a reliable input for subsequent scene grouping.

[0130] Step S43, according to the target scene list, extract the associated feature set of each scene to generate a feature vector.

[0131] Specifically, the target scene list contains multiple possible switching scenes, each scene is associated with a set of features, such as song frequency, interaction intensity, or scene-specific virtual prop usage. The system extracts these features from the interaction data to generate a feature vector for each scene.

[0132] In one possible implementation, for the multi-person chorus scene, the system extracts features including the number of chorus with virtual characters, song frequency, and average dwell time, generating a multi-dimensional feature vector, for example, a vector containing 10 features. Similarly, the feature vector of the virtual stage performance scene may include virtual prop usage frequency and user performance gesture interaction intensity; the feature vector is processed by normalization to ensure that the feature values of different scenes are within the same numerical range, facilitating subsequent clustering analysis.

[0133] In another embodiment, if the virtual KTV platform supports dynamic scene expansion, the system can dynamically generate feature vectors for new scenes, such as virtual DJ mixing. For example, the frequency and dwell time of user adjustments to sound effects in the mixing scene are extracted to generate corresponding feature vectors. This method improves the adaptability of the system to new scenes.

[0134] Step S44, the target scene is grouped using a clustering algorithm to obtain a scene grouping result.

[0135] Specifically, the clustering algorithm is used to group similar target scenes together to provide more accurate switching recommendations for users. The clustering algorithm can be a K-means algorithm or a hierarchical clustering algorithm based on the similarity of scene feature vectors.

[0136] In one embodiment, the system uses a K-means algorithm to divide the scenes in the target scene list into three groups, such as high-interactive scenes, low-interactive scenes, and performance-oriented scenes. During the grouping process, the algorithm calculates the Euclidean distance between each scene feature vector and assigns similar scenes to the same group. For example, multi-person karaoke and virtual DJ mixing scenes are grouped together due to high interaction intensity, while single-person karaoke scenes are grouped separately due to low interaction intensity.

[0137] In another possible implementation, if the scene features are complex, a hierarchical clustering algorithm can be used to generate a tree-like grouping structure. For example, the system first divides all scenes into two categories: interactive and non-interactive, and then further divides the interactive category into karaoke-oriented and mixing-oriented. The grouping result provides a basis for subsequent scene priority ranking.

[0138] Step S45, through the scene grouping result, the priority ranking of each group is obtained to determine the final switching scene sequence.

[0139] Specifically, the scene grouping result reflects the similarity of different scenes, and the system assigns a priority to each group based on the user's current behavior and historical preferences.

[0140] For example, the priority can be determined based on the user's recent karaoke frequency or interaction intensity preferences.

[0141] In one embodiment, if the user recently prefers high-interactive scenes, the system increases the priority of the group containing multi-person karaoke and virtual DJ mixing, and generates a switching sequence containing these two scenes. For example, the sequence is multi-person karaoke to virtual DJ mixing to virtual stage performance; during the ranking process, the system can consider the predicted probability of the scene and the user's preferences through a weighting algorithm, for example, assigning a higher weight to high-interactive scenes.

[0142] In another possible implementation, the system dynamically adjusts the priority according to the user's historical switching records. For example, if the user frequently switches from single-person karaoke to multi-person karaoke, the priority of the multi-person karaoke group is raised, and this method improves the personalization of the switching sequence.

[0143] Step S46, if the priority of the final switching scene sequence meets the preset condition, the scene sequence is output, otherwise the feature vector is re-extracted for clustering.

[0144] Specifically, the preset condition can be that the prediction probability of the highest priority scene in the switching sequence exceeds a threshold value such as 0.7. If the condition is met, the system outputs the final switching scene sequence, for example, multi-person karaoke to virtual DJ mixing. If the condition is not met, it means that the current grouping result may not be accurate, and the system returns to step S43 to re-extract the feature vector and perform clustering.

[0145] In one embodiment, the system checks that the prediction probability of the multi-person karaoke, the highest priority scene in the switching sequence, is 0.75, meeting the threshold requirement, and directly outputs the sequence. If the probability is only 0.6, the system re-extracts the feature vector, for example, adds recent interaction data, and re-executes clustering and sorting. This iterative mechanism ensures the reliability of the switching sequence.

[0146] Step S5, the highest probability target scene is extracted from the possible target scene list, and the change trend of the real-time monitored behavior data is combined to determine whether to switch to the karaoke area, to obtain the accurate switching demand prediction result.

[0147] Specifically, the highest probability target scene is determined by the target scene list generated in step S4, and the system further confirms the switching demand by real-time monitoring of the change trend of the user behavior data. For example, the karaoke area is a high-interactive scene, and the typical features include increased karaoke frequency and multi-person voice interaction.

[0148] In one possible implementation, the system uses a probability sorting algorithm to extract the highest probability scene from the target scene list, for example, the probability of the multi-person karaoke scene is 0.8. The real-time monitoring system continuously collects user behavior data through sensors, for example, recording the karaoke frequency and interaction intensity once every second, to generate a behavior data change trend vector. If the trend vector shows that the karaoke frequency has significantly increased, and the user starts voice interaction with the virtual character, the system confirms that the target scene is the karaoke area and generates a switching signal.

[0149] In another embodiment, if the change trend of the user behavior data does not clearly point to the karaoke area, for example, the karaoke frequency does not increase, the system updates the behavior data through the continuous monitoring module and recalculates the trend vector. This dynamic monitoring mechanism improves the accuracy of switching prediction.

[0150] Step S51, the highest probability target scene is extracted by using the probability sorting algorithm to obtain the highest probability target scene identifier.

[0151] Specifically, the probability sorting algorithm sorts the scenes in the target scene list from high to low according to the prediction probability, and selects the one with the highest probability as the target scene. For example, if the probability of the multi-person chorus scene is 0.8 and the probability of the virtual stage performance is 0.15, the multi-person chorus is the highest probability target scene, and its identifier is generated.

[0152] In one embodiment, the system sorts the scene probabilities by the quicksort algorithm to ensure computational efficiency. After the identifier is generated, the system stores the identifier of the multi-person chorus scene as the input of the subsequent switching signal, which simplifies the scene selection process and improves real-time performance.

[0153] Step S52, collect behavior data through real-time monitoring system, and calculate the change trend of behavior data by using time series analysis method to obtain behavior data change trend vector.

[0154] Specifically, the real-time monitoring system collects user behavior data at a high frequency, such as 10 times per second, including dwell time, song ordering frequency and interaction intensity. The time series analysis method calculates the change trend of the data through a sliding window, for example, the increment of the song ordering frequency in the past 5 seconds.

[0155] In one possible implementation, the system calculates the change rate of behavior data by using the difference method, for example, the song ordering frequency increases from 1 per minute to 3 per minute, and generates a positive change trend vector, which comprehensively reflects the change direction and amplitude of the dwell time, song ordering frequency and interaction intensity, providing a basis for switching judgment.

[0156] In another embodiment, if the virtual KTV scene supports real-time interaction of multiple people, the system can introduce weighted time series analysis, focusing on the change of interaction intensity. For example, the number of chorus between the user and the virtual role increases, and the weight of interaction intensity in the trend vector increases. This method improves the pertinence of trend analysis.

[0157] Step S53, if the behavior data change trend vector points to the preset characteristics of the chorus area, determine the target scene as the chorus area, and generate a scene switching signal.

[0158] Specifically, the preset characteristics of the chorus area include high song ordering frequency, multi-person voice interaction and high interaction intensity. The system compares the similarity between the behavior data change trend vector and the characteristics of the chorus area to determine whether the switching condition is met.

[0159] In an embodiment, if the trend vector shows that the song ordering frequency increases to 3 times per minute, and the user inputs multiple chorus instructions through the microphone, the system confirms that the target scene is a chorus area and generates a switching signal. The switching signal contains the identification and priority of the chorus area, which is used to trigger subsequent resource loading.

[0160] In another possible implementation, if the trend vector partially meets the chorus area characteristics, such as an increase in song ordering frequency but insufficient interaction intensity, the system can determine whether to generate a switching signal through threshold judgment. This method avoids mis-switching and improves the stability of scene switching.

[0161] Step S54, according to the scene switching signal, obtain the configuration parameters related to the chorus area, and determine the switching requirement.

[0162] Specifically, the configuration parameters of the chorus area include the required audio processing module, virtual character interaction logic, and interface rendering requirements of the scene. The system extracts these parameters from the preset configuration library according to the switching signal.

[0163] In an embodiment, the chorus area requires high-fidelity audio processing and multi-person voice synchronization functions, and the system extracts the corresponding audio bandwidth allocation parameters and virtual character animation configuration to generate a switching requirement description. The switching requirement specifies the resource requirements and priority of the chorus area, providing guidance for resource preloading.

[0164] In another possible implementation, if the virtual KTV supports customized chorus scenes, the system can adjust the configuration parameters according to the user's historical preferences. For example, if the user prefers a specific style of virtual character, the system will preferentially load related animation resources. This personalized configuration improves user experience.

[0165] Step S55, if the behavior data change trend vector does not point to the preset characteristics of the chorus area, update the behavior data through the continuous monitoring module to obtain a new behavior data change trend vector.

[0166] Specifically, if the trend vector does not meet the chorus area characteristics, such as no significant increase in song ordering frequency, the system continues to collect behavior data through the continuous monitoring module to update the trend vector.

[0167] In an embodiment, the continuous monitoring module updates the behavior data every second, recalculates the change rates of song ordering frequency and interaction intensity, and generates a new trend vector. For example, if the user starts to frequently order songs after 5 seconds, the new trend vector may meet the chorus area characteristics, triggering the switching signal.

[0168] In another possible implementation, the system can extend the monitoring window, such as from 5 seconds to 10 seconds, to capture longer-term behavior changes. This method is suitable for cases where the user's behavior pattern is more complex.

[0169] Step S56, the trend vector of the new behavior data is classified by using the random forest algorithm to determine whether the chorus area switching condition is met, and a classification result is obtained.

[0170] Specifically, the random forest algorithm classifies the trend vector through multiple decision trees to determine whether the chorus area feature is met. The input of the algorithm is the trend vector, and the output is the classification result, which is met or not met.

[0171] In an embodiment, the random forest contains 100 decision trees, and each tree is trained based on a subset of the trend vector. For example, one tree analyzes the change of song ordering frequency, and another tree analyzes the interaction intensity, and the classification result is generated by comprehensive voting. If 80% of the decision trees determine that the chorus area feature is met, the system confirms that the switching condition is established.

[0172] In another possible implementation, the system can optimize the random forest model through feature importance analysis. For example, if the song ordering frequency contributes the most to the classification result, the system will monitor this feature first to improve the classification efficiency. This method improves the accuracy of switching judgment.

[0173] Step S57, according to the classification result, a final switching demand prediction result is determined, and a target scene switching instruction is generated.

[0174] Specifically, the classification result determines whether to generate a switching instruction. If the chorus area feature is met, the system generates a switching instruction containing the identification of the chorus area; otherwise, the classification model is continuously monitored or adjusted.

[0175] In an embodiment, if the random forest algorithm confirms that the trend vector meets the chorus area feature, the system generates a switching instruction, which contains the resource loading priority and interface adjustment parameters of the chorus area. This instruction ensures the smoothness of the switching process and improves the user's immersion in the virtual KTV.

[0176] So far, the method realizes the complete process from user behavior analysis to target scene determination through scene switching prediction, classification and priority sorting, which provides support for subsequent resource preloading and interface optimization.

[0177] It should be noted that in this embodiment, the multi-scene data refers to a comprehensive data set related to user behavior and scene attributes covering multiple different scenes in the virtual environment of the song ordering station, which is used to mine cross-scene association features and support accurate prediction of scene switching.

[0178] The feedback data refers to the closed-loop data extracted from the seamless interface state after the completion of the scene switching process, which is used to verify the prediction accuracy and optimize the model. The purpose of the feedback data is to realize the iterative upgrade of the long short-term memory network.

[0179] The trigger timing refers to a trigger timing of the preloading mechanism, that is, a specific time node of a resource file preloading operation of a system starting a target scene, such as audio, image, interactive logic component and interface element; the essence is to solve the contradiction between too early resource preloading leading to bandwidth, cache redundancy and too late leading to scene switching delay, finally support the core goal of seamless interface connection, ensure that the resources of the target scene have been loaded when the user actually performs the switching operation, and realize the scene transition without lag.

[0180] In the description of the present application, it should be understood that the terms "coaxial", "bottom", "one end", "top", "middle", "the other end", "upper", "one side", "top", "inner", "front", "central", "both ends" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the present application.

[0181] In addition, the terms "first", "second", "third", "fourth" are only for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated, so that the features with "first", "second", "third", "fourth" can be explicitly or implicitly included at least one of the features.

[0182] In the present application, unless otherwise explicitly specified and limited, the terms "mounting", "setting", "connecting", "fixing", "threading" and the like should be understood broadly, for example, it can be fixedly connected, or it can be detachably connected, or it can be integrated; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium; it can be the internal communication of two elements or the interaction relationship between two elements, unless otherwise explicitly limited, the above-mentioned terms in the present application can be understood according to the specific meaning of the above-mentioned terms in the present application by those skilled in the art.

[0183] Although the embodiments of the present application have been shown and described, it can be understood by those skilled in the art that various changes, modifications, replacements and variations can be made to the embodiments without departing from the principles and spirits of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A multi-scene switching method based on a point song station, characterized in that, The application comprises the following steps: Collecting behavior data of users in a virtual environment through sensors, including stay time, song ordering frequency and interaction intensity, and processing the behavior data by using a long short-term memory network to obtain a user dynamic preference vector; Matching the user dynamic preference vector with a preset scene type library, which includes at least one of a single-person song ordering scene, a multi-person karaoke scene and a virtual stage performance scene, and if the matching degree is higher than a threshold, it is determined that the current scene preference is stable, otherwise it is determined that a potential switching intention exists, and the potential switching intention probability distribution is obtained through the above determination; After obtaining the potential switching intention probability distribution, the support vector machine classifier is used to train the correlation characteristics between multiple scene data to obtain a possible target scene list for cross-scene switching; Extracting the target scene with the highest probability from the possible target scene list, and combining the real-time monitored behavior data change trend to determine whether the behavior data change trend points to a karaoke area, and if so, it is determined that the karaoke area will be switched to, otherwise the monitoring continues to obtain an exact switching demand prediction result; According to the exact switching demand prediction result, the resource files and interface elements of the target scene are preloaded, the dynamic allocation algorithm is used to adjust the server bandwidth allocation ratio to obtain an optimized resource configuration scheme; Updating the interface layout parameters through the optimized resource configuration scheme, if the prediction result matches the actual user operation, the layout parameters are directly applied, if the prediction result does not match the actual user operation, the layout parameters are rolled back to the original configuration, and the final interface seamless connection state is determined; Extracting feedback data in the final interface seamless connection state and inputting it into the iterative training process of the long short-term memory network to update the model parameters to adapt to the user behavior mode.

2. The method of claim 1, wherein the method further comprises: The step of collecting behavior data of users in a virtual environment through sensors, including stay time, song ordering frequency and interaction intensity, and processing the behavior data by using a long short-term memory network to obtain a user dynamic preference vector, comprises the following steps: Collecting user behavior data from a virtual environment through sensors, including stay time, song ordering frequency and interaction intensity, and storing them as structured time series data to obtain an initial behavior data set; Using data cleaning technology to pre-process the initial behavior data set, removing noise data and missing values, and aligning the stay time, song ordering frequency and interaction intensity through timestamps to obtain standardized behavior sequences; According to the long short-term memory network, the standardized behavior sequences are time-series modeled to extract time-dependent features and generate a user dynamic preference vector; According to the user dynamic preference vector, a clustering algorithm is used to group users based on the similarity of behavior mode vectors to obtain user preference categories; If the user preference category matches the preset music recommendation rule, a corresponding song list is obtained from the music library according to the category features to generate a personalized recommendation sequence; According to the comparison between the personalized recommendation sequence and the user song ordering frequency, the matching degree of the recommendation sequence and the actual behavior is calculated to obtain a recommendation effect evaluation index; According to the recommendation effect evaluation index, the long short-term memory network parameters are optimized by using the gradient descent algorithm to update the user dynamic preference vector.

3. The method of claim 1, wherein the method further comprises: The user dynamic preference vector is matched with a preset scene type library, the scene type library includes at least one of a single person song ordering scene, a multi-person karaoke scene, and a virtual stage performance scene, if the matching degree is higher than a threshold, it is judged that the current scene preference is stable, otherwise it is judged that the potential switching intention exists, and the potential switching intention probability distribution step is obtained by the above judgment, including: Obtain the user dynamic preference vector, generate a feature vector by analyzing the user historical interaction data; Based on the vector matching algorithm, compare the user dynamic preference vector with the standard vector in the preset scene type library to obtain the matching score; If the matching score is higher than the preset threshold, it is judged that the current scene preference is stable, and a stable scene label is generated; According to the stable scene label, extract the recommended content related to the current scene to generate a personalized recommendation list; If the matching score is lower than the preset threshold, analyze the potential switching intention by the Bayesian inference algorithm to obtain the intention probability distribution; According to the intention probability distribution, a weighted sorting algorithm is used to generate a recommended sequence of potential switching scenes; According to the recommended sequence, dynamically adjust the user interaction interface, and output the optimized scene recommendation content.

4. The method of claim 1, wherein the method further comprises: After obtaining the potential switching intention probability distribution, the support vector machine classifier is used to train the correlation characteristics between multi-scene data to obtain the possible target scene list of cross-scene switching, including: Obtain the user interaction data, and extract the probability distribution of the potential switching intention from the data; Through probability distribution analysis, the support vector machine classifier is used to train the correlation characteristics of multi-scene data to obtain a classification model; If the prediction accuracy of the classification model reaches a preset threshold, the target scene list of cross-scene switching is determined according to the model output.

5. The method of claim 1, wherein: The steps of extracting the highest probability target scene from the possible target scene list and combining the real-time monitored behavior data change trend judgment, if the behavior data change trend points to the karaoke area, it is determined that the karaoke area will be switched to, otherwise the monitoring is continued to obtain the exact switching demand prediction result, including: Obtain the target scene from the scene list, extract the highest probability target scene by using the probability sorting algorithm to obtain the highest probability target scene identifier; Collect behavior data through a real-time monitoring system, calculate the change trend of the behavior data by using a time series analysis method to obtain a behavior data change trend vector; If the behavior data change trend vector points to the preset characteristics of the karaoke area, the target scene is determined as the karaoke area, and a scene switching signal is generated; According to the scene switching signal, obtain the configuration parameters related to the karaoke area, and determine the switching demand; If the behavior data change trend vector does not point to the preset characteristics of the karaoke area, update the behavior data through a continuous monitoring module to obtain a new behavior data change trend vector; Classify the new behavior data change trend vector by using a random forest algorithm to determine whether the karaoke area switching condition is met, and obtain a classification result; According to the classification result, determine the final switching demand prediction result, and generate a target scene switching instruction.

6. The method of claim 1, wherein: The step of preloading resource files and interface elements of the target scene according to the prediction result of the exact switching demand, adopting a dynamic allocation algorithm to adjust the server bandwidth allocation ratio, and obtaining an optimized resource configuration scheme, comprises: Through the prediction model, analyze the user behavior data to obtain the prediction result of the switching demand; Based on the prediction result, determine the priority and resource demand of the target scene, and generate a resource loading sequence; According to the resource loading sequence, pre-load the resource files and interface elements of the target scene and store them in the local cache; Using a consistent hash algorithm, dynamically allocate server bandwidth and adjust the bandwidth ratio of each scene; If the loading speed of a scene is lower than the preset threshold after adjusting the bandwidth ratio, then through the load balancing mechanism, re-allocate the bandwidth to obtain an optimized bandwidth allocation scheme; According to the optimized bandwidth allocation scheme, monitor the resource loading state in real time to obtain the loading completion degree data; According to the loading completion degree data, dynamically adjust the triggering time of the preloading mechanism to generate the final resource configuration scheme.

7. The method of claim 1, wherein the method further comprises: The step of updating the interface layout parameters through the optimized resource configuration scheme, and if the prediction result matches the actual user operation, directly applying the layout parameters, otherwise rolling back to the original configuration to determine the final seamless interface switching state, comprises: By analyzing the resource configuration scheme, obtain the currently available computing resources and storage resources from the preset resource pool to determine the real-time allocation state of the resource configuration scheme; According to the real-time allocation state of the resource configuration scheme, generate the corresponding interface layout parameters using the preset interface rendering algorithm to obtain the preliminarily adjusted layout parameters; Through the machine learning classification model, predict the preliminarily adjusted layout parameters to obtain the prediction result and determine whether the prediction result conforms to the expected user operation mode; If the prediction result is consistent with the actual user operation, then through the parameter application logic, deploy the preliminarily adjusted layout parameters to the interface rendering module to determine the update state of the interface layout; If the prediction result is inconsistent with the actual user operation, then from the original layout parameter storage library, obtain the original layout parameters and use the rollback operation mechanism to restore the interface to the initial state to obtain the rolled-back interface state; According to the rolled-back interface state or the update state, use the state synchronization algorithm to detect the delay and consistency in the interface switching process to determine the seamless switching state; By recording the seamless switching state and user operation data, update the training data set of the dynamic adjustment mechanism to obtain the optimized resource configuration scheme and obtain the input parameters for the next adjustment.

8. A multi-scene switching system based on a jukebox, characterized by, Comprise: The behavior data processing module is configured to collect the user's stay time, song ordering frequency and interaction intensity behavior data in the virtual environment through the sensor, process the behavior data using a long short-term memory network to obtain a user dynamic preference vector; The scene preference judgment module is configured to match the user dynamic preference vector with a preset scene type library, the scene type library including at least one of a single-person song ordering scene, a multi-person karaoke scene, and a virtual stage performance scene, and if a matching degree is higher than a threshold, it is judged that a current scene preference is stable, otherwise, it is judged that a potential switching intention exists, and a potential switching intention probability distribution is obtained through the above judgment; The associated feature training module is configured to, after obtaining the potential switching intention probability distribution, train associated features between multi-scene data by using a support vector machine classifier to obtain a possible target scene list of cross-scene switching; The switching demand prediction module is configured to extract a target scene with the highest probability from the possible target scene list, and judge in combination with a real-time monitored behavior data change trend, if the behavior data change trend points to a karaoke area, it is determined that the karaoke area will be switched to, otherwise, the monitoring is continued to obtain an exact switching demand prediction result; The resource configuration optimization module is configured to, according to the exact switching demand prediction result, pre-load resource files and interface elements of the target scene, adjust a server bandwidth allocation ratio by using a dynamic allocation algorithm, and obtain an optimized resource configuration scheme; The interface connection determination module is configured to update interface layout parameters by using the optimized resource configuration scheme, if the prediction result matches an actual user operation, the layout parameters are directly applied, if the prediction result does not match the actual user operation, the layout parameters are rolled back to original configurations, and a final interface seamless connection state is determined; The model iterative updating module is configured to extract feedback data in the final interface seamless connection state, input the feedback data into an iterative training process of the long short-term memory network, and update model parameters to adapt to a user behavior mode.

9. A song ordering apparatus comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein, The processor executes the computer program to implement the steps of the multi-scene switching method based on a song ordering table according to any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the multi-scene switching method based on a song ordering table according to any one of claims 1 to 7.

Images