Response optimization method, system, device and medium based on a jukebox

Abstract

The application belongs to the field of human-computer interaction, and specifically discloses a response optimization method, system, device and medium based on a song ordering machine, wherein the method comprises: acquiring behavior sequence features by collecting user click play and skip behavior data on the song ordering machine, combining time period and activity scene information; obtaining a user behavior dynamic change index after processing the behavior sequence features by using behavior sequence time sequence coding; analyzing time correlation patterns in the behavior sequence by using a sequence model according to the user behavior dynamic change index; obtaining a behavior change trend through a time pattern extraction mechanism and dynamic index quantitative calculation; and if the behavior change trend shows that the skip rate of a certain type of song is higher than a preset threshold, corresponding time period and scene factors are extracted. The application aims to solve the problem of high song skip rate caused by short-term boredom or scene inadaptability of users in the song ordering scene in the prior art.

Description

Technical Field

[0001] This invention relates to the field of human-computer interaction technology, specifically to a response optimization method, system, device, and medium based on a karaoke machine. Background Technology

[0002] In the context of the rapid development of digital entertainment, karaoke machines, as an important carrier of music consumption, directly impact user experience and platform stickiness through the performance of their intelligent recommendation systems. Karaoke machine recommendation systems analyze user behavior to provide personalized song rankings, aiming to improve user satisfaction. However, with the diversification of user needs and changes in scenarios, recommendation systems need to possess dynamic adaptability to ensure that the recommendations highly match users' real-time preferences. The core of this field lies in how to capture subtle changes in user behavior through technological means and optimize the accuracy and timeliness of song ranking accordingly, thereby increasing user acceptance of recommended content. Current common practices in karaoke machine recommendation systems rely on preset rating factor weights, such as song popularity, user preferences, or historical playback records, to calculate recommendation rankings using fixed formulas. However, these methods often fall short when faced with rapidly changing user behavior patterns. Especially at different times or in different scenarios, users' preferences for song types may change significantly; for example, users may prefer light music on weekday mornings and pop songs on weekend nights. Existing static weight settings struggle to capture such dynamic changes, leading to a disconnect between recommendation results and actual user needs, thus impacting user experience.

[0003] A deeper technical challenge lies in accurately monitoring the dynamic changes in user behavior and adjusting weighting coefficients accordingly. User clicks, play completion rates, or skipping actions on recommended songs reflect their preferences, but these behaviors often conceal complex contextual information. For example, a user's music choices at different times may be influenced by emotions, activity scenarios, or social environments. Existing systems typically lack in-depth analysis of this contextual information when processing this behavioral data, resulting in untargeted weighting adjustments that fail to accurately reflect subtle differences in user needs. Furthermore, even if changes in user behavior can be monitored, translating these changes into reasonable adjustments to weighting coefficients remains a critical challenge. The diversity and non-linearity of user behavior make a single weighting adjustment method unsuitable for all scenarios. For example, a user frequently skipping a certain type of song in a short period might indicate temporary aversion to that type of music, but it could also be due to the current context being unsuitable for that type of music. If the system simply reduces the weight of that type of song, it may misjudge the user's long-term preferences, causing subsequent recommendations to deviate from the user's true needs. Therefore, how to accurately analyze the contextual information of user behavior in dynamically changing scenarios and adaptively adjust the weight coefficients accordingly has become a key issue for karaoke machine recommendation systems to improve ranking performance. Summary of the Invention

[0004] This invention provides a response optimization method, system, device, and medium based on karaoke machines, aiming to solve the problem of high song skipping rates in karaoke machine scenarios due to short-term user boredom or discomfort with the environment.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0006] A response optimization method based on karaoke machines includes: acquiring behavioral sequence features by collecting user click-play and skip behavior data on the karaoke machine, combined with time period and activity scenario information; processing the behavioral sequence features using behavioral sequence temporal coding to obtain dynamic change indicators of user behavior; analyzing time-related patterns in the behavioral sequence using a sequence model based on the dynamic change indicators of user behavior, and obtaining behavioral change trends through a time pattern extraction mechanism and dynamic indicator quantification; if the behavioral change trend shows that the skip rate for a certain type of song is higher than a preset threshold, extracting the corresponding time period and scenario factors, and using skip rate threshold comparison and context parsing rule definition to determine whether the change stems from short-term boredom or scenario incompatibility to obtain context influence weights; updating the song type preference coefficient in the fixed weights based on the context influence weights and trend prediction error correction behavioral change rate evaluation, and using pattern correlation... The analysis and fusion of indicators and weight adjustments yield an adaptive weight set. The comprehensive score of songs is calculated using this adaptive weight set, and the songs are sorted and optimized based on the current user's real-time preferences. This incorporates time-period factor decomposition and scenario incompatibility index quantification to obtain a preliminary recommendation list. If the matching degree between the songs in the preliminary recommendation list and historical playback records is lower than a preset threshold, a reinforcement learning model is used to fine-tune the song order in the preliminary recommendation list. Short-term boredom pattern recognition and influence weight calculation formulas are used to determine whether the adjusted list improves the matching degree, resulting in an optimized recommendation sequence. Subsequent user interaction feedback data is obtained based on the optimized recommendation sequence, combined with factor extraction logic chains and judgment basis data fusion to obtain the proportion of positive behavior in the feedback data. Based on the proportion of positive behavior in the feedback data, the parameters of the sequence model are updated, and after optimization of the sequence model parameters, a behavior change prediction model for the next cycle is obtained.

[0007] In one aspect of this disclosure, the step of obtaining dynamic change indicators of user behavior by collecting user click-play and skip behavior data on the karaoke machine, combining time period and activity scenario information to obtain behavioral sequence features, and processing the behavioral sequence features using behavioral sequence temporal coding, includes:

[0008] By collecting user click play and skip behavior data on the karaoke machine, and combining it with time period and activity scenario, user behavior sequences are generated;

[0009] The behavior sequence is processed using a time window segmentation method to extract sequence features and obtain the temporal distribution features of the behavior sequence;

[0010] If the frequency of clicks to play in the time distribution features is higher than a preset threshold, then a long short-term memory network is used to perform temporal encoding on the behavior sequence to obtain the encoded behavior sequence vector.

[0011] Based on the encoded behavior sequence vector, the similarity of user behavior patterns in different time periods and activity scenarios is calculated to obtain behavior pattern clustering results.

[0012] Based on the clustering results of the aforementioned behavioral patterns, the dynamic change trends of users in specific activity scenarios are extracted to obtain dynamic change indicators.

[0013] Based on the aforementioned dynamic change indicators, combined with time period and scenario association, a user behavior preference sequence is generated to obtain a user behavior prediction model;

[0014] According to one aspect of this disclosure, the steps of analyzing time-related patterns in a behavior sequence using a sequence model based on the user behavior dynamic change indicators, and obtaining the behavior change trend through a time pattern extraction mechanism and dynamic indicator quantification calculation, include:

[0015] User behavior sequence data is acquired, and time series decomposition method is used to extract time-related patterns and obtain the time features of the behavior sequence.

[0016] By using the time series features and employing a long short-term memory network model, the dynamic change patterns in the behavioral sequence are analyzed to obtain behavioral pattern prediction results.

[0017] If the behavior pattern prediction results deviate significantly from the preset threshold, the sliding window method is used to calculate the dynamic index quantification value to obtain the quantitative characteristics of behavior changes.

[0018] Based on the quantitative characteristics of the behavioral changes, a decision tree model is used to determine the category of the behavioral change trend and obtain the trend classification result.

[0019] Based on the trend classification results and combined with the time-related pattern, a weighted average method is used to calculate the intensity of the behavioral change trend and obtain the trend intensity value.

[0020] The trend strength value is judged. If the strength value exceeds a preset threshold, the stability of the behavior change trend is obtained through clustering method, thereby obtaining trend stability characteristics.

[0021] Based on the aforementioned trend stability characteristics, a time series smoothing method is used to optimize the output of the behavioral change trend and obtain the final behavioral change trend.

[0022] In one aspect of this disclosure, the step of extracting the corresponding time period and scenario factors, and using skip rate threshold comparison and context parsing rules to determine whether the change stems from short-term boredom or scenario discomfort to obtain context influence weights if the behavioral change trend shows that the skip rate for a certain type of song is higher than a preset threshold, includes:

[0023] Based on the skip rate data of various songs obtained by users, statistical analysis is used to calculate the mean and variance of the skip rate and obtain the trend of skip rate changes.

[0024] If the skip rate trend is higher than the preset threshold, the context parsing module will extract the scene factors and time period factors during user interaction to generate a context feature set.

[0025] The context feature set is classified using a decision tree algorithm to determine whether the change in skip rate stems from short-term boredom or scene discomfort, and the classification result is obtained.

[0026] Based on the classification results, the contextual impact weights of short-term boredom and scene discomfort are calculated, and a weighted average method is used to generate a comprehensive impact weight.

[0027] The recommendation priority of song categories is adjusted by the comprehensive influence weights to generate an updated recommendation sequence;

[0028] Based on the updated recommendation sequence, the A / B testing module is used to verify the recommendation effect, obtain user interaction data and the optimized skip rate;

[0029] If the optimized skip rate is still higher than the preset threshold, the interaction data is input into the context parsing module through the feedback loop module to regenerate the context feature set.

[0030] In one aspect of this disclosure, the step of evaluating and updating the song type preference coefficient in the fixed weights based on the context influence weights combined with the rate of change of trend prediction error correction behavior, and obtaining an adaptive weight set through pattern correlation analysis and index fusion weight adjustment, includes:

[0031] Historical data is obtained from user interaction records, and user preferences are extracted using time series analysis to obtain a set of user preference features;

[0032] Based on the user preference feature set, the K-means clustering algorithm is used to divide the song types and obtain song type groups.

[0033] By grouping by song type, the rate of behavioral change is calculated to obtain the trend of behavioral change.

[0034] Based on the trend of behavioral change, if the rate of change exceeds a preset threshold, a linear regression algorithm is used to predict the trend deviation and obtain the trend prediction error.

[0035] By analyzing the trend prediction error and combining it with pattern correlation analysis, the fixed weights are adjusted to obtain the weight update coefficients.

[0036] Based on the weight update coefficients and the fusion index fusion method, the adaptive weight set is updated to obtain the optimized weight set;

[0037] By optimizing the weight set and reallocating the recommendation priority of song types, the final recommendation ranking is obtained.

[0038] In one aspect of this disclosure, the step of calculating the comprehensive score of songs through the adaptive weight set, optimizing the ranking based on the current user's real-time preferences, incorporating time-time factor decomposition and scene incompatibility index quantification, and obtaining a preliminary recommendation list includes:

[0039] Real-time preference data is obtained based on users' historical interaction data and current behavior, and user preference vectors are obtained through collaborative filtering algorithms.

[0040] Based on the real-time preference vector and song metadata, an adaptive weight set is calculated to obtain the weight allocation for each song;

[0041] By analyzing the characteristics of the current time period through a time period segmentation model, and combining this with user preference vectors, the weight set is adjusted to generate a time-weighted score.

[0042] A scene discomfort index quantitative model is used to evaluate the matching degree between the song and the current scene, and the scene discomfort value is calculated.

[0043] The overall score of a song is calculated by combining time-weighted scores and scene-inappropriateness values, which forms the basis for ranking.

[0044] Songs are sorted in descending order based on their overall scores, and a preliminary recommendation list is generated using a quick sorting algorithm.

[0045] We update real-time preference data based on user feedback, adjust the adaptive weight set, and optimize the recommendation list.

[0046] In one aspect of this disclosure, the step of obtaining the proportion of positive behavior in the feedback data by combining the user's subsequent interaction feedback data obtained from the optimized recommendation sequence with the factor extraction logic chain and judgment basis data fusion includes:

[0047] We acquire user interaction feedback data based on recommendation sequences, and use time series analysis to extract timestamps and interaction types of user behavior to obtain a user behavior dataset.

[0048] Based on the user behavior dataset, cluster analysis is used to divide the behavior into categories, determine the classification boundaries between positive and negative behaviors, and obtain the behavior classification results.

[0049] Based on the positive behavior weight distribution, the correlation of each behavior node in the logical chain is extracted, and the correlation strength between nodes is analyzed by Pearson correlation coefficient to obtain the behavior correlation matrix;

[0050] The proportion of positive behaviors in the recommendation sequence is calculated by using the behavior association matrix.

[0051] If the proportion of positive behavior is lower than the preset threshold, the parameters of the recommended sequence are adjusted, user interaction feedback data is reacquired, and an updated behavior dataset is obtained.

[0052] Based on the updated behavior dataset, repeat the behavior classification and proportion determination steps to obtain the final positive behavior proportion.

[0053] In another aspect, this disclosure also relates to a response optimization system based on a karaoke machine, comprising:

[0054] The behavior sequence feature acquisition module is configured to acquire behavior sequence features based on the collected user click play and skip behavior data on the karaoke machine, combined with time period and activity scenario information, and then obtain the user behavior dynamic change index by processing the behavior sequence features with behavior sequence temporal coding.

[0055] The user behavior dynamic change index generation module is configured to analyze the time-related patterns in the behavior sequence using a sequence model based on the user behavior dynamic change index, and obtain the behavior change trend through a time pattern extraction mechanism and dynamic index quantification calculation.

[0056] The behavior change trend determination module is configured to extract the corresponding time period and scene factors if the behavior change trend shows that the skip rate of a certain type of song is higher than a preset threshold, and use the skip rate threshold comparison and context parsing rule definition to determine whether the change is due to short-term boredom or scene discomfort to obtain the context influence weight.

[0057] The context influence weight acquisition module is configured to update the song type preference coefficient in the fixed weight based on the context influence weight and the rate of change of trend prediction error correction behavior, and to obtain an adaptive weight set through pattern correlation analysis and index fusion weight adjustment.

[0058] The adaptive weight set generation module is configured to calculate the comprehensive score of the songs through the adaptive weight set, optimize the sorting based on the current user's real-time preferences, incorporate time period factor decomposition and scene incompatibility index quantification, and obtain a preliminary recommendation list.

[0059] The initial recommendation list determination module is configured to, if the matching degree between the songs in the initial recommendation list and the historical playback records is lower than a preset threshold, use a reinforcement learning model to fine-tune the order of the songs in the initial recommendation list, and use short-term boredom pattern recognition and influence weight calculation formula to determine whether the adjusted list improves the matching degree, thereby obtaining an optimized recommendation sequence.

[0060] The optimized recommendation sequence acquisition module is configured to obtain the proportion of positive behavior in the feedback data by combining the user's subsequent interaction feedback data with the factor extraction logic chain and judgment basis data fusion based on the optimized recommendation sequence.

[0061] The behavior change prediction model update module is configured to update the parameters of the sequence model based on the proportion of positive behavior in the feedback data, and incorporate the optimized sequence model parameters to obtain the behavior change prediction model for the next period.

[0062] In another aspect of this disclosure, there is also a karaoke device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described response optimization method based on a karaoke machine.

[0063] In another aspect, this disclosure also relates to a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described response optimization method based on a karaoke machine.

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

[0065] This invention collects user click-to-play and skip behavior data, combines it with time period and activity scenario information to generate behavioral sequence features, and uses time-series coding and sequence models to analyze behavioral change trends, accurately identifying the causes of abnormal skip rates. The invention extracts scenario factors through a context parsing module, calculates context influence weights, and dynamically adjusts song type preference coefficients by combining trend prediction error correction and behavioral change rate evaluation to form an adaptive weight set and optimize the recommendation list ranking. If the recommendation list has a low match with historical playback, reinforcement learning is used to fine-tune the ranking, incorporating short-term boredom pattern recognition to improve recommendation accuracy. Finally, user feedback data is used to optimize sequence model parameters, forming the prediction model for the next cycle. This invention significantly improves the personalized adaptation capability of recommendation systems, effectively reduces skip rates caused by scenario incompatibility or short-term boredom, and improves user experience and interaction satisfaction. Attached Figure Description

[0066] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0067] Figure 1 This is a flowchart of the karaoke machine response optimization method in this invention.

[0068] Figure 2 This is one of the schematic diagrams of the karaoke machine response optimization method in this invention.

[0069] Figure 3 This is the second schematic diagram of the karaoke machine response optimization method in this invention. Detailed Implementation

[0070] The present invention will be further described below with reference to embodiments. These embodiments are merely some, not all, of the embodiments of the present invention. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are all within the protection scope of the present invention.

[0071] Example 1

[0072] Please see Figures 1-3As shown, this embodiment discloses a response optimization method based on a karaoke machine, including: acquiring behavioral sequence features based on collected user click-play and skip behavior data on the karaoke machine, combined with time period and activity scenario information; processing the behavioral sequence features using behavioral sequence time-series encoding to obtain dynamic change indicators of user behavior; analyzing time-related patterns in the behavioral sequence using a sequence model based on the dynamic change indicators of user behavior, and obtaining behavioral change trends through a time pattern extraction mechanism and dynamic indicator quantification; if the behavioral change trend shows that the skip rate for a certain type of song is higher than a preset threshold, extracting the corresponding time period and scenario factors, and using skip rate threshold comparison and context parsing rule definition to determine whether the change originates from short-term boredom or scenario discomfort to obtain context influence weights; updating the song type preference coefficient in the fixed weights based on the context influence weights combined with trend prediction error correction behavioral change rate evaluation, and... An adaptive weight set is obtained through pattern correlation analysis and indicator fusion weight adjustment. The comprehensive score of songs is calculated using this adaptive weight set, and the songs are sorted and optimized based on the current user's real-time preferences. This incorporates time-period factor decomposition and scenario incompatibility index quantification to obtain a preliminary recommendation list. If the matching degree between the songs in the preliminary recommendation list and historical playback records is lower than a preset threshold, a reinforcement learning model is used to fine-tune the order of the songs in the preliminary recommendation list. Short-term boredom pattern recognition and influence weight calculation formulas are used to determine whether the adjusted list improves the matching degree, resulting in an optimized recommendation sequence. Subsequent user interaction feedback data is obtained based on the optimized recommendation sequence, combined with factor extraction logic chains and judgment basis data fusion to obtain the proportion of positive behavior in the feedback data. The parameters of the sequence model are updated based on the proportion of positive behavior in the feedback data, and after optimization of the sequence model parameters, a behavior change prediction model for the next cycle is obtained.

[0073] This invention collects user click-to-play and skip behavior data, combines it with time period and activity scenario information to generate behavioral sequence features, and uses time-series coding and sequence models to analyze behavioral change trends, accurately identifying the causes of abnormal skip rates. The invention extracts scenario factors through a context parsing module, calculates context influence weights, and dynamically adjusts song type preference coefficients by combining trend prediction error correction and behavioral change rate evaluation to form an adaptive weight set and optimize the recommendation list ranking. If the recommendation list has a low match with historical playback, reinforcement learning is used to fine-tune the ranking, incorporating short-term boredom pattern recognition to improve recommendation accuracy. Finally, user feedback data is used to optimize sequence model parameters, forming the prediction model for the next cycle. This invention significantly improves the personalized adaptation capability of recommendation systems, effectively reduces skip rates caused by scenario incompatibility or short-term boredom, and improves user experience and interaction satisfaction.

[0074] Example 2

[0075] Please see Figures 1-3As shown, this embodiment is a further optimization based on Embodiment 1. In this embodiment, the present invention provides a personalized recommendation method based on user behavior, applicable to karaoke machine scenarios. By analyzing user interaction behavior, time periods, and scenario factors, it generates an accurate song recommendation list. The method is described in detail below with reference to specific implementation methods.

[0076] Step S1: Obtain user click play and skip behavior data on the karaoke machine, and generate user behavior sequence by combining time period and activity scenario information.

[0077] Specifically, when users use the karaoke machine, they will click to play or skip via the touch screen or buttons, and this behavioral data can be recorded through the karaoke machine's interaction log.

[0078] Time period information refers to the time range in which user operations occur, such as being divided into morning, afternoon, evening, or even more specifically, hourly.

[0079] The activity scenario information includes the environment in which the user uses the karaoke machine, such as family gatherings, bar entertainment, or personal leisure. These scenarios can be obtained through the karaoke machine's built-in sensors, manual selection by the user, or presets in the background.

[0080] For example, in a family gathering scenario, users might prefer upbeat pop songs, while in a bar scenario, they might prefer dance music with a strong beat. By combining click-to-play and skip behavior data with time periods and scenario information, a behavior sequence is formed that includes timestamps, behavior types, and scenario tags. For example, a behavior sequence might be recorded as: a user clicked to play a pop song at 8 pm in a family gathering scenario, and then skipped a rock song 5 seconds later.

[0081] In one possible implementation, the generation of the behavior sequence can be achieved by collecting data in real time through the karaoke machine's log system and arranging it in chronological order.

[0082] The logging system records the occurrence time, behavior type, and scene tag for each action, ensuring data integrity. To improve data quality, the collected behavioral data can be preprocessed, such as removing duplicate clicks or invalid operations.

[0083] The preprocessing process includes checking the continuity of timestamps to ensure that the behavior sequence reflects the user's actual interaction habits. It should be noted that the behavior sequence includes not only clicks to play and skip, but also other interactions such as pause and favorite, to enrich the sequence information.

[0084] For example, in a family gathering scenario, a user clicked to play 5 songs and skipped 3 songs between 7 pm and 9 pm. The system recorded these behaviors and labeled the time period as "evening" and the scenario as "family gathering." In this way, behavioral sequences can comprehensively capture users' preferences at specific times and in specific scenarios, providing a foundation for subsequent analysis.

[0085] Step S2: The behavior sequence is processed using the time window segmentation method to extract sequence features and obtain the time distribution features of the behavior sequence.

[0086] Specifically, the time window segmentation method divides a behavioral sequence into multiple subsequences according to fixed or dynamic time intervals to analyze the distribution pattern of user behavior over time. The size of the time window can be adjusted according to the actual scenario, for example, set to 5 minutes, 10 minutes, or 1 hour.

[0087] In a family gathering scenario, considering that users may operate frequently in a short period of time, the time window can be set to a shorter 5 minutes to capture rapidly changing behavioral patterns. Each segmented time window contains several behavioral data, such as the number of times the user clicks to play, the number of times the user skips, and the corresponding scene tags.

[0088] By analyzing behavioral data within each time window, sequence features are extracted, including the frequency of clicks to play, the frequency of skipping, and the distribution of behavioral types.

[0089] In one embodiment, the extraction of temporal distribution features can be achieved through statistical analysis. For example, for each time window, the percentage of clicks to play and skips, as well as the average interval between these actions, can be calculated. These features reflect the user's behavioral tendencies within a specific time period. For instance, if the frequency of clicks to play is much higher than the frequency of skips within a certain time window, it may indicate that the user is satisfied with the currently recommended songs; conversely, if the frequency of skips is high, it may reflect the user's dissatisfaction with the recommended content. It should be noted that temporal distribution features can also include the periodicity of behavioral sequences, such as users being more inclined to choose a certain type of song on weekend evenings and a different type of song on weekday mornings.

[0090] For example, in a bar scenario, the system segments the behavior sequence into 10-minute time windows. A user, within the time window of 10 PM to 10:10 PM, clicked to play 4 dance tracks and skipped 2 slow songs. Through statistical analysis, the system extracted the characteristics of this time window: the click frequency was 0.4 times / minute, the skip frequency was 0.2 times / minute, and the playback ratio of dance tracks was higher than that of slow songs. These temporal distribution characteristics provided data support for subsequent time-series coding.

[0091] Step S21: If the frequency of clicks to play in the time distribution features is higher than a preset threshold, then use a long short-term memory network to perform temporal encoding on the behavior sequence to obtain the encoded behavior sequence vector.

[0092] Specifically, the frequency of clicks to play songs reflects the user's acceptance of the recommended songs. A preset threshold can be set based on historical data or scenario requirements. For example, in a family gathering scenario, the threshold can be set to 0.3 times / minute, indicating that frequent song selection suggests high satisfaction with the recommended content. When the click frequency exceeds this threshold, it indicates that the behavioral sequence has a strong temporal regularity, making it suitable for encoding using a Long Short-Term Memory (LSTM) network. An LTM network is a neural network model capable of capturing long-term dependencies in time series, generating a fixed-dimensional behavioral sequence vector by processing the temporal distribution characteristics of the behavioral sequence.

[0093] In one possible implementation, the input to the Long Short-Term Memory (LSTM) network is a sequence of behavioral features within a time window, including click counts, skip counts, and scene labels. The network processes these features through multiple layers of neurons, extracting temporal patterns from the sequence, such as the trend of a user's preference for a particular type of song within a continuous time window. The encoded behavioral sequence vector is a multi-dimensional vector containing dynamic information about user behavior, such as the rhythm of preference changes or behavioral patterns in specific scenes. It should be noted that the network's training data can come from a large number of user behavioral sequences, and the training process optimizes the loss function to enable the network to accurately capture the temporal patterns of the behavioral sequences.

[0094] For example, in a personal leisure scenario, the user clicks to play 0.5 times per minute, which is higher than the preset threshold of 0.3 times per minute.

[0095] The system inputs the behavioral sequence of this time window into a Long Short-Term Memory (LSTM) network. The network analyzes the temporal distribution characteristics of the sequence and outputs a 64-dimensional behavioral sequence vector. This vector captures the user's persistent preference for popular music and lower interest in slow songs in this scenario. This encoding method allows subsequent behavioral pattern analysis to be based on more compact and information-rich data.

[0096] Step S22: Based on the encoded behavior sequence vector, calculate the similarity of user behavior patterns in different time periods and activity scenarios to obtain behavior pattern clustering results.

[0097] Specifically, behavioral pattern similarity measures whether user behavior sequences share similar characteristics across different time periods and scenarios. Similarity calculation can be performed using methods such as cosine similarity or Euclidean distance. The encoded behavioral sequence vector contains temporal information about user behavior. By comparing vectors from different time periods or scenarios, the system can identify commonalities in user behavior. For example, in a family gathering scenario, users might tend to choose popular songs at both 8 PM and 9 PM; this similarity can be reflected through vector comparison.

[0098] The clustering process uses a density-based clustering method to group similar behavioral patterns into one category, generating behavioral pattern clustering results.

[0099] In one embodiment, the system first collects behavioral sequence vectors across multiple time periods and scenarios, such as vectors from a user's weekend family gatherings and weekday personal leisure activities. Next, it calculates the similarity between these vectors to generate a similarity matrix. Based on the similarity matrix, a clustering algorithm is used to categorize behavioral patterns into several classes, such as frequently playing popular songs and occasionally skipping slow songs.

[0100] The clustering results reflect the behavioral patterns of users in different scenarios, providing a basis for subsequent extraction of dynamic change trends.

[0101] It should be noted that scene labels can be introduced as additional constraints during the clustering process to ensure that the clustering results are relevant to the actual scene.

[0102] For example, in a bar scenario, the system analyzes the behavioral sequence vectors of users on Friday night and Saturday night and finds that the two have a high similarity in the pattern of clicking to play dance music, with a cosine similarity of 0.9.

[0103] The system categorizes these vectors into the dance music preference category, while classifying the user's behavior vectors in the morning personal leisure scenario into the light music preference category. This clustering result can clearly distinguish the user's behavior patterns in different scenarios.

[0104] Step S23: Extract the dynamic change trend of users in specific activity scenarios through the behavior pattern clustering results to obtain dynamic change indicators.

[0105] Specifically, the behavior pattern clustering results provide a classification of user behavior in different scenarios, while the dynamic change trend further analyzes the changing patterns of these classifications over time.

[0106] For example, users may initially prefer pop songs in a family gathering setting, but gradually shift towards dance music over time; the extraction of dynamic change trends can be achieved by analyzing the changes in behavioral patterns over time in the clustering results; the dynamic change index is a quantitative value that reflects the magnitude and direction of changes in user behavioral preferences, such as the rate of change from a preference for pop songs to a preference for dance music.

[0107] In one possible implementation, the system performs time-series analysis on the behavioral sequence vectors within each cluster, calculating the changing trend of behavioral patterns within consecutive time windows. For example, by comparing the vector differences between adjacent time windows, the switching frequency and intensity of behavioral patterns can be determined; dynamic change indicators can include information such as the rate of change, the direction of change, and the stability of the change. For instance, if a user's vectors in a family gathering scenario show an increased preference for dance music across three consecutive time windows, the system can extract dynamic change indicators of this increased dance music preference.

[0108] For example, in a family gathering scenario, the system analysis of clustering results revealed that users' behavior pattern was to frequently play popular songs between 7 pm and 8 pm, while gradually shifting to frequently playing dance music between 8 pm and 9 pm.

[0109] By calculating the vector change rate, the system obtains a dynamic change index: dance music preference is enhanced, with a change rate of 0.2 / hour. This index reflects the gradual increase in users' interest in dance music in this scenario, providing data support for the subsequent generation of preference sequences.

[0110] Step S3: Using dynamically changing indicators and combining time periods and scenarios, a user behavior preference sequence is generated to obtain a user behavior prediction model.

[0111] Specifically, dynamic change indicators provide patterns of user behavior changes over time and context. By combining time period and context information, the system further generates user behavior preference sequences. These preference sequences are ordered sequences that record user preference categories in different time periods and contexts, such as 8 PM, family gatherings, and popular song preferences. The process of generating preference sequences involves associating dynamic change indicators with time periods and context labels to determine user preference patterns under specific conditions. The user behavior prediction model is trained using machine learning methods based on these preference sequences and is used to predict users' future behavioral tendencies.

[0112] In one embodiment, the system first integrates dynamic change indicators with time periods and scene labels to form sequence data containing time, scene, and preference information.

[0113] For example, a preference sequence might be: 7 PM, family gathering, pop music preference; 8 PM, family gathering, dance music preference. The system then uses this sequence data to train a prediction model, such as one based on a recurrent neural network. During training, the model learns the temporal and contextual relationships within the preference sequence; for example, a user is more likely to choose dance music after 8 PM in a family gathering setting. After training, the model can predict the user's preference category based on the input time period and context.

[0114] For example, in a bar setting, the system generates preference sequences based on dynamically changing metrics: dance music preference at 10 PM; electronic music preference at 11 PM. The system uses this sequence data to train a recurrent neural network model. The model learns that users in a bar setting are more likely to choose upbeat music over time. After training, the model can predict the user's preferences at the next moment, given the time and setting; for example, predicting that the user might choose electronic music at 11:30 PM. This predictive ability improves the accuracy of recommendations and avoids recommending songs that the user is not interested in.

[0115] In one possible implementation, the generation of preference sequences can be supplemented by historical user data. For example, the system analyzes a user's behavioral sequence in family gathering scenarios over the past month, discovering that users typically choose popular songs on weekend evenings and prefer light music on weekday evenings. By combining dynamic change indicators, the system generates a more comprehensive preference sequence, covering various time and scenario combinations. The trained predictive model can then generate more user-friendly recommended content based on the current time and scenario.

[0116] It's important to note that the training process for the predictive model needs to be updated regularly to adapt to changes in user preferences. For example, the system can collect new behavioral data weekly and retrain the model to ensure that the predictions are consistent with the user's latest behavior. This dynamic updating approach improves the model's adaptability and maintains the stability of the recommendation performance.

[0117] Step S4: Generate a personalized recommendation list based on the user behavior prediction model to obtain the recommended content for the karaoke machine.

[0118] Specifically, the user behavior prediction model analyzes user behavior preference sequences to predict a user's song preferences in a specific time period and scenario. The personalized recommendation list is generated based on the prediction model's output, combined with the karaoke machine's song library, to filter out songs that match the user's preferences. The process of generating the recommendation list involves matching the predicted preference categories with song features in the song library; for example, associating a preference for popular songs with songs tagged as popular music in the library. The recommendation list is sorted by matching degree, ensuring that songs the user is most likely to be interested in are listed first.

[0119] In one embodiment, the system first obtains preference prediction results for the current time period and scenario from a user behavior prediction model, such as predicting that the user prefers popular songs at 8 pm in a family gathering scenario. Next, the system extracts all popular songs from the song library and calculates the matching degree with the user's preferences based on the song's metadata, such as artist, release year, and rhythm. The matching degree can be calculated using vector similarity, for example, by comparing the song's feature vector with the user's preference vector. Finally, the system generates a recommendation list containing 10 songs, arranged in descending order of matching degree.

[0120] For example, in a bar scenario, the predictive model outputs that the user prefers dance music at 10 PM. The system selects 100 dance tracks from its music library and generates a recommendation list containing the top 10 tracks by comparing the similarity of song features, such as rhythm and style, with the user's preference vector. This recommendation list can meet the user's music needs in the current scenario, improving the user experience.

[0121] In one possible implementation, the generation of the recommendation list can incorporate diversity constraints to avoid recommending overly homogenous songs. For example, when filtering popular songs, the system can ensure that the list includes songs by different artists or with different tempos, thereby increasing the richness of the recommended content. Diversity constraints can be implemented by setting a threshold for feature differences between recommended songs, such as requiring that the tempo difference between adjacent songs in the list be greater than a certain value.

[0122] It's worth noting that the generation of recommendation lists can also incorporate users' long-term preferences. For example, the system might analyze a user's playback history during family gatherings over the past month and discover a preference for a particular artist's songs. When generating the recommendation list, the system can prioritize popular songs by that artist, thereby further improving the accuracy of the recommendations. This approach, combining long-term and short-term preferences, maintains the stability of recommendations even in dynamically changing scenarios.

[0123] Step S5: Obtain user behavior sequence data, use time series decomposition method to extract time-related patterns, and obtain the time features of the behavior sequence.

[0124] Specifically, user behavior sequence data includes user interaction records such as clicking play and skipping on a karaoke machine. Combined with timestamps and context information, time series decomposition methods are used to analyze the patterns of behavior sequences over time. For example, these can be decomposed into components such as trends, periodicity, and random fluctuations. The trend component reflects the overall direction of user behavior changes over time, such as a shift in preference from pop songs to dance music; the periodic component captures recurring patterns of behavior within specific time periods, such as a user's preference for dance music on Friday nights; and random fluctuations represent unpredictable short-term behavioral changes. The temporal characteristics are a comprehensive representation of these components, used to describe the temporal patterns of user behavior.

[0125] In one embodiment, the system performs time-series decomposition on the behavioral sequence to extract trend and periodic components. For example, the system analyzes a user's behavioral sequence over the past week and finds that users click to play music more frequently on weekend evenings than on weekdays, and that the preference for dance music gradually increases. The decomposition process can be implemented using the moving average method, first calculating the trend component of the behavioral sequence, and then extracting the periodic component through Fourier transform.

[0126] Ultimately, the system generates time features, including trend slope and periodic frequency, for subsequent dynamic change pattern analysis.

[0127] For example, in a family gathering scenario, the system analyzes users' behavior sequence over the past month, decomposing it into trend components: users' preference for popular songs gradually increases between 7 PM and 9 PM; the periodic component shows that users click to play songs more frequently on weekend evenings. Temporal features include a trend slope of 0.15 / hour and a periodicity frequency of once a week. These features provide a foundation for analyzing the dynamic changes in user behavior.

[0128] In one possible implementation, time series decomposition can be combined with contextual information. For example, in a bar setting, the system might detect a strong periodicity in user behavior after 10 PM, manifested as clicking to play a dance track every 20 minutes. By decomposing and extracting these patterns, the system forms temporal features, providing more accurate data support for subsequent model analysis.

[0129] Step S51: Using time series features and a long short-term memory network model, analyze the dynamic change patterns in the behavior sequence to obtain behavior pattern prediction results.

[0130] Specifically, time-series features contain trend and periodic information about user behavior, making them suitable for analysis using Long Short-Term Memory (LSTM) network models. LTM networks, by capturing long-term and short-term dependencies in time series data, can identify dynamic patterns of user behavior change, such as a shift from a preference for slow music to a preference for dance music. The behavioral pattern prediction output of the network represents the user's likely behavioral tendencies in the future, such as predicting that a user is more likely to choose electronic music in the next hour.

[0131] In one embodiment, the system inputs time-series features into a Long Short-Term Memory (LSTM) network. The network processes these features through multiple layers of neurons to extract dynamic patterns from the behavioral sequences. For example, the network analyzes the time characteristics of users during family gatherings and discovers that users gradually increase their playback of dance music after 8 PM. The network outputs a prediction: the user has an 80% probability of choosing dance music in the next time window. During training, the network uses historical behavioral sequence data and optimizes the loss function to make the predictions more closely approximate actual behavior.

[0132] For example, in a bar setting, the system inputs the user's time-series characteristics from 9 PM to 10 PM into a Long Short-Term Memory (LSTM) network. These characteristics include trends and periodic patterns in play frequency. After analysis, the network predicts that the user is likely to continue preferring dance music between 10 PM and 11 PM, and that the play frequency will increase to 0.5 times per minute. This prediction provides a basis for quantifying subsequent behavioral changes.

[0133] It's important to note that training Long Short-Term Memory (LSTM) networks can incorporate data from various scenarios. For example, the system can collect behavioral sequences from family gatherings, bars, and individual leisure settings to train a general-purpose model. The model can then adjust its predictions based on the input scenario labels, adapting to behavioral patterns in different scenarios. This cross-scenario training approach enhances the model's generalization ability.

[0134] Step S52: If the behavior pattern prediction result deviates significantly from the preset threshold, the sliding window method is used to calculate the dynamic index quantification value to obtain the quantitative characteristics of behavior change.

[0135] Specifically, the deviation between the predicted behavior pattern and the actual behavior reflects the accuracy of the prediction model. A preset threshold can be the difference between the predicted probability and the actual behavior, for example, set to 0.2. If the deviation exceeds this threshold, it indicates that the prediction may be inaccurate, and further analysis of the quantitative characteristics of behavioral changes is needed. The sliding window method calculates dynamic indicators within each window, such as the rate of behavioral change or the magnitude of frequency change, by sliding a fixed-size window across the behavioral sequence. These indicators form quantitative characteristics that describe the dynamic patterns of user behavior changes.

[0136] In one possible implementation, the size of the sliding window can be adjusted according to the scenario, for example, set to 10 minutes in a family gathering scenario. Within each window, the system calculates the rate of change in the frequency of clicks to play and skip, for example, increasing from 0.3 clicks / minute to 0.5 clicks / minute. Quantitative features include information such as the rate of change, the direction of change, and the magnitude of change. For example, if the user increases the skip frequency over three consecutive windows, the system generates the quantitative feature: the skip frequency change rate is 0.1 clicks / minute / window. These features provide data support for subsequent trend classification.

[0137] For example, in a personal leisure scenario, the system predicts that the user prefers light music, but actual behavior shows that the user frequently skips light music, with a deviation of 0.3, which is higher than the threshold of 0.2. The system uses a sliding window method to analyze the behavior sequence over the past 30 minutes, calculating the change in skipping frequency for each 10-minute window, and obtains a quantitative feature: the rate of change in skipping frequency is 0.15 times / minute / window. This feature indicates that the user's interest in light music is declining, providing a basis for adjusting the recommendation strategy.

[0138] Step S53: Based on the quantitative characteristics of behavioral changes, a decision tree model is used to determine the category of behavioral change trends and obtain trend classification results.

[0139] Specifically, quantitative features provide a numerical representation of changes in user behavior. Decision tree models analyze these features to determine the category of behavioral change trends, such as preference enhancement, preference weakening, or preference stabilization. Based on the magnitude and distribution of feature values, the decision tree model constructs classification rules; for example, if the skipping frequency changes at a rate greater than 0.1 times / minute / window, it is classified as preference weakening. The trend classification result is the model's output, representing the type of user behavior change.

[0140] In one embodiment, the system inputs quantified features into a decision tree model, which then classifies the data according to preset rules.

[0141] For example, the rules could include: if the rate of change of the click-to-play frequency is greater than 0.2 times / minute / window, it is classified as preference enhancement; if the rate of change of the skip frequency is greater than 0.1 times / minute / window, it is classified as preference weakening.

[0142] The system analyzes the quantitative characteristics of users in family gathering scenarios and obtains trend classification results: users' preference for dance music has increased. This result provides a basis for subsequent trend strength calculation.

[0143] For example, in a bar scenario, the system analyzes quantitative features and finds that the rate of change in the frequency of users clicking to play electronic music is 0.25 times / minute / window, which is higher than the threshold of 0.2. Based on this feature, the decision tree model classifies the user as having an enhanced preference.

[0144] This classification result indicates that users' interest in electronic music is increasing, and the system can adjust the recommendation list accordingly to increase the proportion of electronic music.

[0145] It's important to note that decision tree models can be trained based on extensive historical user behavior data. For example, the system can collect quantitative characteristics of users in different scenarios to train the model and identify common trend categories. During training, the model optimizes classification accuracy to ensure the reliability of the classification results. This approach improves the model's adaptability to different scenarios.

[0146] Step S54: Based on the trend classification results and combined with the time-related pattern, the intensity of the behavioral change trend is calculated using a weighted average method to obtain the trend intensity value.

[0147] Specifically, the trend classification results provide the categories of behavioral changes, while the time-related patterns contain trend and periodic information of the behavioral sequence. The weighted average method calculates the strength of the behavioral change trend by assigning weights to the quantitative features and the time-related patterns. For example, the weight of the rate of change of click-to-play frequency can be set to 0.6, and the weight of the periodic frequency can be 0.4. The trend strength value is a numerical value that represents the significance of the behavioral change. For example, a higher value indicates a stronger change in user preferences.

[0148] In one possible implementation, the system combines the trend classification results with time-related patterns to calculate a weighted average. For example, in a family gathering scenario, the classification result is an enhanced preference, and the time-related pattern shows that users click to play music more frequently after 8 PM. The system assigns a weight of 0.7 to the quantitative feature and a weight of 0.3 to the time-related pattern, calculating a trend strength value of 0.85; this value indicates that the change in users' preference for dance music is relatively significant.

[0149] For example, in a bar scenario, the system analyzes the trend classification results "enhanced preference" and time-related patterns, finding that users play dance music more frequently after 10 PM. After weighted averaging, the trend strength value is 0.9, higher than the threshold of 0.8, indicating a strong change in user preference for dance music. Based on this, the system can prioritize recommending dance music to improve user satisfaction.

[0150] Step S55: Based on the trend strength value, if the strength value exceeds a preset threshold, the stability of the behavior change trend is determined by clustering method to obtain trend stability characteristics.

[0151] Specifically, the trend strength value reflects the significance of behavioral changes. If the strength value exceeds a preset threshold, such as 0.8, it indicates a significant trend, requiring further analysis of its stability. Clustering methods analyze trend strength values ​​across multiple time windows to categorize similar trends and determine whether the trend is stable. For example, if the trend strength values ​​across multiple time windows differ little, the trend is stable; if the differences are large, the trend fluctuates significantly. Trend stability is a quantitative indicator representing the reliability of the changing trend.

[0152] In one embodiment, the system collects trend intensity values ​​over 10 consecutive time windows and uses a clustering method to classify these values ​​into two categories: stable and fluctuating.

[0153] For example, in a family gathering scenario, the system analyzes the trend intensity values ​​of 10 time windows and finds that the intensity values ​​of 8 windows are between 0.8 and 0.9, and the clustering result is stable. The stable trend characteristic indicates that the user's preference for dance music is continuous, and the system can confidently increase the proportion of dance music recommendations.

[0154] For example, in a personal leisure scenario, the system found that the trend strength value fluctuated between 0.7 and 0.95, and the clustering result was "fluctuation". The stable trend feature indicates that the user's preference for light music is not stable enough. The system can retain other types of songs when making recommendations to avoid the recommendations being too monotonous.

[0155] Step S56: Using the trend stability feature, the time series smoothing method is used to optimize the output of the behavioral change trend, and the final behavioral change trend is obtained.

[0156] Specifically, trend stability features provide reliable information about behavioral change trends. Time series smoothing methods smooth trend strength values, reducing the interference of short-term fluctuations and generating a more stable trend output. For example, smoothing methods can use exponential moving averages to weight trend strength values ​​across consecutive time windows, generating a smoothed trend curve. The final behavioral change trend is a continuous trend representation used to guide adjustments to recommendation strategies.

[0157] In one possible implementation, the system smooths the intensity value sequence in the trend stability feature. For example, in a family gathering scenario, the system analyzes the trend intensity values ​​over 10 time windows, uses an exponential moving average method to generate a smoothed trend curve, showing that users' preference for dance music is continuously increasing. This trend provides a reliable basis for subsequent recommendations.

[0158] For example, in a bar scenario, the system smoothed the trend strength values ​​and found that users' preference for electronic music steadily increased after 10 PM. The final behavioral trend indicated that the system should increase the proportion of electronic music recommendations during this time period to improve user acceptance of the recommended content.

[0159] Step S6: Obtain skip rate data for various types of songs, and use statistical analysis to calculate the mean and variance of the skip rate to obtain the trend of skip rate changes.

[0160] Specifically, skip rate data records users' skipping behavior for different types of songs, such as the proportion of users skipping popular songs. Statistical analysis analyzes the distribution patterns of skipping behavior by calculating the mean and variance of the skip rate. The mean reflects the average level of skipping behavior, while the variance represents the volatility of skipping behavior. Skip rate trends are determined by comparing the mean skip rate over consecutive time windows to assess changes in user acceptance of a particular type of song.

[0161] In one embodiment, the system collects skip rate data for popular songs and dance music during family gatherings. Statistical analysis shows that the mean skip rate for popular songs is 0.2, with a variance of 0.05; the mean skip rate for dance music is 0.1, with a variance of 0.03. The skip rate trend indicates that users have a higher acceptance of dance music than popular songs, and the skipping behavior is more stable.

[0162] For example, in a bar scenario, the system analyzes users' skip rate for electronic music and finds a mean of 0.15 and a variance of 0.04, with the skip rate gradually decreasing after 10 PM. This skip rate trend indicates that users' acceptance of electronic music increases during this time period, allowing the system to adjust its recommendation strategy accordingly.

[0163] Step S61: If the skip rate trend is higher than the preset threshold, the context parsing module extracts the scene factors and time period factors during user interaction to generate a context feature set.

[0164] Specifically, if the skip rate trend exceeds a preset threshold, such as 0.3, it indicates a significant decrease in user acceptance of a certain type of song, requiring analysis of the contextual factors leading to skipping. The context parsing module extracts scene information from user interactions, such as family gatherings, bars, and time periods, such as 8 PM, to generate a contextual feature set. The feature set includes scene tags, timestamps, and statistical features of interaction behavior, such as skip frequency and playback duration.

[0165] In one possible implementation, the system analyzes user skipping behavior in a family gathering scenario and finds that the skipping rate reaches 0.4 after 9 PM, exceeding the threshold of 0.3. The context parsing module extracts the following features: scenario: family gathering; time: 9 PM; skipping frequency: 0.4 times / minute. The generated context feature set provides data support for subsequent classification analysis.

[0166] For example, in a personal leisure scenario, the system found that the user's skip rate for slow songs reached 0.35 at 8 AM, exceeding the threshold of 0.3. The context parsing module extracted the following features: scenario: personal leisure; time: 8 AM; skip frequency: 0.35 times / minute. The context feature set indicates that the user has lower interest in slow songs during this time period, and the system can reduce the proportion of slow songs recommended.

[0167] Step S62: Use the decision tree algorithm to classify the context feature set, determine whether the change in skip rate is due to short-term boredom or scene incompatibility, and obtain the classification result.

[0168] Specifically, the contextual feature set contains scene and time information. The decision tree algorithm analyzes these features to determine the reasons for changes in the skip rate. Classification rules can include: if the skip frequency increases rapidly in a short period of time, it is classified as short-term boredom; if the skipping behavior is highly related to a specific scene, it is classified as scene incompatibility. The classification results indicate the main driving factors of the skipping behavior, providing a basis for subsequent weight adjustments.

[0169] In one embodiment, the system inputs a set of contextual features into a decision tree model, which classifies the content according to rules. For example, in a family gathering scenario, if the skipping frequency increases from 0.1 times / minute to 0.4 times / minute within 10 minutes, the decision tree model classifies it as "short-term boredom." This result suggests that the user may be experiencing fatigue with the currently recommended song type.

[0170] For example, in a bar setting, the system analyzes the contextual feature set and finds that users' skipping of slow songs mainly occurs during high-intensity music sessions after 10 PM. The decision tree model classifies this as "scene inappropriate," indicating that slow songs do not match the current context. Based on this, the system can reduce the recommendation ratio of slow songs and increase dance music or electronic music.

[0171] Step S63: Calculate the contextual impact weights of short-term boredom and scene discomfort based on the classification results, and generate the comprehensive impact weights using a weighted average method.

[0172] Specifically, the classification results indicate the reasons for skipping behavior, while contextual influence weights are used to quantify the contribution of short-term boredom and contextual discomfort to the skipping behavior. Weight calculations can be based on the confidence level of the classification results; for example, short-term boredom has a confidence level of 0.7, and contextual discomfort has a confidence level of 0.3. A weighted average method generates a comprehensive influence weight by assigning weights to the two factors, which is used to adjust the recommendation strategy.

[0173] In one possible implementation, the system calculates a weight of 0.6 for short-term boredom, a weight of 0.4 for scene incompatibility, and a combined impact weight of 0.64 based on the classification results. This weight indicates that short-term boredom contributes more to skipping behavior, and the system can prioritize adjusting song types to reduce recommendations of songs that the user is already tired of.

[0174] For example, in a family gathering scenario, the classification results show that skipping behavior is mainly due to short-term boredom, with a weight of 0.8, while scenario incompatibility has a weight of 0.2. The overall impact weight is 0.76. Based on this, the system adjusts the recommendation list, reducing the proportion of popular songs and increasing the proportion of dance music to improve user satisfaction.

[0175] Step S64: Adjust the recommendation priority of song categories by comprehensively considering the influence weights, and generate an updated recommendation sequence.

[0176] Specifically, the overall influence weight reflects the reason for skipping behavior, and the system adjusts the recommendation priority of song categories based on the weight. For example, if the weight of "short-term boredom" is high, the system reduces the priority of song types that the user is already tired of; if the weight of "inappropriate scenario" is high, the system reduces song types that do not match the current scenario. The updated recommendation sequence is reordered according to the adjusted priority to ensure that the recommended content better meets the user's needs.

[0177] In one embodiment, the system lowers the recommendation priority of popular songs and increases the priority of dance music based on a comprehensive influence weight of 0.76. The generated new recommendation sequence contains more dance songs and ranks them higher. This adjustment can reduce user skipping behavior and improve recommendation effectiveness.

[0178] For example, in a bar setting, the overall impact weighting indicates that inappropriate scene conditions are the primary cause. The system lowers the recommendation priority of slow songs and increases the priority of electronic music. The new recommendation sequence contains 10 electronic music songs, sorted in descending order of relevance, resulting in a significant decrease in user skip rate.

[0179] Step S65: For the updated recommendation sequence, use the A / B testing module to verify the recommendation effect, obtain user interaction data, and obtain the optimized skip rate.

[0180] Specifically, the A / B testing module verifies the effectiveness of the new recommendation sequence by comparing it with the original recommendation sequence. The system divides users into two groups: one group receives the updated recommendation sequence, and the other receives the original sequence. Interaction data for both groups is recorded, such as click-through rate and skip rate. The optimized skip rate, representing the skip rate of the new sequence, reflects the degree of improvement in recommendation effectiveness.

[0181] In one possible implementation, the system conducts A / B testing in a family gathering scenario, with the updated recommended sequence containing more dance music. Test results show that the skip rate of the new sequence decreased from 0.4 to 0.2, indicating a significant improvement in recommendation performance. User interaction data supports subsequent feedback loops.

[0182] For example, in a bar scenario, A / B testing showed that the skip rate for the updated recommended sequence, such as one mainly featuring electronic music, was 0.15, lower than the original sequence's 0.3. The optimized skip rate indicates that the new sequence better meets user needs, and the system can use this sequence as the default recommended content.

[0183] In step S66, if the optimized skip rate is still higher than the preset threshold, the interaction data is input into the context parsing module through the feedback loop module to regenerate the context feature set.

[0184] Specifically, if the optimized skip rate is still higher than a preset threshold, such as 0.3, it indicates that the recommended sequence still needs further adjustment. The feedback loop module inputs user interaction data from A / B testing, such as skip behavior and timestamps, into the context parsing module, re-analyzes the scenario and time factors, and generates a new set of contextual features. This process forms a closed-loop optimization, ensuring that the recommended sequence is continuously improved.

[0185] In one embodiment, the system found that the optimized skip rate was 0.35, which is higher than the threshold of 0.3. The feedback loop module inputs the interaction data into the context parsing module to extract new features: the scenario is a family gathering, the time is 10 pm, and the skip frequency is 0.35 times / minute. The new context feature set provides data support for subsequent classification and weight adjustment.

[0186] For example, in a personal leisure scenario, the optimized skip rate is 0.4, which is higher than the threshold of 0.3. The feedback loop module re-inputs the interaction data into the context parsing module to generate a new feature set, showing that the user's skipping behavior for slow songs increases at 8 AM. The system adjusts its recommendation strategy based on the new feature set, reducing the proportion of slow songs recommended.

[0187] Step S7: Obtain user's historical interaction records, use time series analysis to extract user preferences, and obtain user preference feature set.

[0188] Specifically, user interaction history records include user actions such as clicking play, skipping, and adding to favorites on the karaoke machine over a past period. Time series analysis extracts user preferences by analyzing the temporal patterns of these records, such as preferences for certain types of songs or artists. The user preference feature set is a multidimensional representation that includes information such as preference category, preference intensity, and temporal distribution.

[0189] In one possible implementation, the system analyzes the user's interaction records over the past month and finds that the user prefers popular songs in family gatherings, with a preference strength of 0.7. Time series analysis extracts the distribution pattern of preferences over time, for example, the user's preference for popular songs is stronger on weekend evenings. The user preference feature set provides the basis for subsequent song type grouping.

[0190] For example, in a bar scenario, the system analyzes a user's interaction history over the past two weeks to extract preference features: the user's preference strength for electronic music is 0.8, and this preference is more pronounced after 10 PM. The feature set includes preference categories such as electronic music, an intensity of 0.8, and a time distribution after 10 PM, providing support for recommendation optimization.

[0191] Step S71: Based on the user preference feature set, use the K-means clustering algorithm to classify song types and determine the song type grouping.

[0192] Specifically, the user preference feature set contains information about users' preferences for different song genres. The K-means clustering algorithm analyzes these features to classify the songs in the music library into several genres, such as pop, dance, and electronic music. The grouping results are based on the similarity between song features, such as rhythm and style, and user preference features, ensuring that the grouping conforms to user habits.

[0193] In one embodiment, the system inputs the user preference feature set and the feature vector of the song library into a K-means clustering algorithm, setting the number of clusters to 5. The algorithm categorizes the songs into five types: pop, dance, electronic music, ballads, and rock. The grouping results show that users have a stronger preference for pop and dance music in family gathering scenarios.

[0194] For example, in a bar setting, the K-means clustering algorithm categorizes songs into three types: electronic music, dance music, and ballads. The grouping results indicate that users' preferences for electronic music and dance music are dominant, and the system can prioritize recommending these two types of songs.

[0195] Step S72: Group by song type, calculate the rate of change of behavior, and obtain the trend of behavior change.

[0196] Specifically, song type grouping provides a basis for classifying songs, while the rate of behavioral change is calculated by analyzing user interactions with each type of song, such as the frequency of clicks to play and skip over time. For example, if a user's frequency of clicking to play dance music increases, the system calculates the rate of change and generates a behavioral change trend.

[0197] In one possible implementation, the system analyzes user interaction behavior with dance music in a family gathering scenario, finding that the click-to-play frequency increased from 0.3 times / minute to 0.5 times / minute, a rate of change of 0.2 times / minute / hour. This behavioral trend indicates an increased user preference for dance music.

[0198] For example, in a bar setting, the system calculates the rate of change in the frequency of user clicks on electronic music, which is 0.25 times per minute per hour. This behavioral trend indicates a continuously increasing user interest in electronic music, allowing the system to optimize recommended content accordingly.

[0199] Step S73: Based on the trend of behavioral change, if the rate of change exceeds a preset threshold, a linear regression algorithm is used to predict the trend deviation and obtain the trend prediction error.

[0200] Specifically, behavioral change trends reflect dynamic changes in user preferences. If the rate of change exceeds a preset threshold, such as 0.2 times / minute / hour, it indicates a significant change, and its future trend needs to be predicted. Linear regression algorithms predict changes over future time periods by fitting a trend line to the rate of behavioral change, and calculate the deviation between the predicted and actual values ​​to obtain the trend prediction error.

[0201] In one embodiment, the system analyzes the rate of change in user behavior regarding dance music and finds that the rate of 0.25 times / minute / hour is higher than the threshold of 0.2. A linear regression algorithm predicts a click-through rate of 0.6 times / minute for the next hour, while the actual value is 0.55 times / minute, resulting in a trend prediction error of 0.05. This error provides a basis for weight adjustment.

[0202] For example, in a personal leisure scenario, the rate of behavioral change (0.3 times / minute / hour) exceeds the threshold of 0.2. Linear regression predicts that the user's playback frequency of light music will decrease, with an error of 0.04. The system adjusts its recommendation strategy based on this error, reducing the proportion of light music recommended.

[0203] Step S74: By analyzing the trend prediction error and combining it with pattern correlation analysis, the fixed weights are adjusted to obtain the weight update coefficients.

[0204] Specifically, trend prediction error reflects the accuracy of the prediction model, while pattern correlation analysis determines the direction of weight adjustment by calculating the correlation between behavioral change trends and user preference characteristics. Fixed weights are the initial recommendation weights assigned to song types, and weight update coefficients are used to dynamically adjust these weights.

[0205] In one possible implementation, the system analyzes a trend prediction error of 0.05 and finds a high correlation between user preferences for dance music and behavioral change trends, with a correlation coefficient of 0.9. The system adjusts the fixed weight of the dance music from 0.5 to 0.7, generating a weight update coefficient of 0.2. This coefficient increases the recommendation priority of the dance music.

[0206] For example, in a bar scenario, the trend prediction error was 0.06, and pattern correlation analysis showed a strong correlation between electronic music and user behavior. The system adjusted the weight of electronic music from 0.4 to 0.65, with a weight update coefficient of 0.25, thus optimizing the recommendation effect.

[0207] Step S75: Update the adaptive weight set according to the weight update coefficients and the fusion index fusion method to obtain the optimized weight set.

[0208] Specifically, the weight update coefficients provide the direction for adjustment, and the indicator fusion method updates the adaptive weight set by comprehensively considering the rate of change of behavior, trend prediction error, and pattern correlation. The optimized weight set is a multi-dimensional vector assigned to different song types to generate the final recommendation ranking.

[0209] In one embodiment, the system updates the adaptive weight set by combining the rate of change of behavior and the results of correlation analysis with a weight update coefficient of 0.2. The optimized weight set shows that dance music has a weight of 0.7 and popular songs have a weight of 0.5. The system adjusts the recommendation priority based on this weight set.

[0210] For example, in a bar scenario, the system combines a weight update coefficient of 0.25 and the rate of behavioral change to generate an optimized weight set: electronic music 0.65, dance music 0.55. The optimized weight set improves the accuracy of recommendations and reduces skipping behavior.

[0211] Step S76: By optimizing the weight set, the recommendation priority of song type is redistributed to obtain the final recommendation ranking.

[0212] Specifically, the optimized weight set assigns new weights to song types, and the system reorders the songs in the song library according to these weights to generate the final recommendation ranking. The ranking process takes into account user preferences, context, and time factors to ensure that the recommended content meets user needs.

[0213] In one possible implementation, the system selects high-weight songs from the song library based on an optimized weight set (0.7 for dance music and 0.5 for pop songs) to generate the final recommendation ranking. The ranking result contains 10 songs, with 6 dance songs and 4 pop songs.

[0214] For example, in a bar scenario, the optimization weight set is 0.65 for electronic music and 0.55 for dance music. The final recommendation ranking includes 8 electronic music tracks and 2 dance music tracks, sorted in descending order of weight. User interaction data shows a significant reduction in skip rate.

[0215] Step S8: Obtain real-time preference data through user historical interaction data and current behavior, and determine the user preference vector using a collaborative filtering algorithm.

[0216] Specifically, user historical interaction data contains long-term preference information, while current behavior reflects short-term preference changes. Collaborative filtering algorithms generate user preference vectors by analyzing the similarity between users and other users, representing the strength of a user's preference for different song types.

[0217] In one embodiment, the system analyzes users' historical interaction data in family gathering scenarios, discovers that users prefer popular songs, and combines this with current behavior, such as frequently playing dance music, to generate a preference vector: popular songs 0.6, dance music 0.8. This vector provides support for subsequent weight calculations.

[0218] For example, in a bar scenario, a collaborative filtering algorithm analyzes a user's playback history with other users and discovers that the user prefers electronic music. Combined with the current behavior, a preference vector is generated: electronic music 0.7, dance music 0.5. This vector improves the personalization of recommendations.

[0219] Step S81: Calculate the adaptive weight set based on the real-time preference vector and song metadata to obtain the weight allocation for each song.

[0220] Specifically, the real-time preference vector represents the user's current preference tendency, and the song metadata includes information such as style and rhythm. The adaptive weight set assigns weights to each song by calculating the similarity between the preference vector and the song feature vector.

[0221] In one possible implementation, the system compares real-time preference vectors with song metadata, calculates cosine similarity, and generates an adaptive set of weights. For example, in a family gathering scenario, dance songs would have a weight of 0.75, while pop songs would have a weight of 0.55.

[0222] For example, in a bar scenario, the adaptive weight set shows that electronic music songs have a weight of 0.8, and dance music has a weight of 0.6. This weight allocation improves the relevance of the recommended content.

[0223] Step S82: Analyze the characteristics of the current time period through the time period segmentation model, combine with the user preference vector, adjust the weight set, and generate a time-weighted score.

[0224] Specifically, the time-segmentation model analyzes the characteristics of the current time period, such as behavioral patterns at 8 PM. Combined with user preference vectors, an adaptive weight set is adjusted to generate a time-weighted score that reflects the match between the song and the current time.

[0225] In one embodiment, the system analyzes a family gathering scenario at 8 PM and discovers that users prefer dance music. A time-weighted score is used to increase the weight of dance music to 0.8, generating a ranking criterion.

[0226] For example, in a bar scenario, time-weighted scores show that electronic music has a weight of 0.85 after 10 pm, which optimizes the recommendation ranking.

[0227] Step S83: Use the scene discomfort index quantitative model to evaluate the matching degree between the song and the current scene, and calculate the scene discomfort value.

[0228] Specifically, the scene discomfort index quantification model calculates the scene discomfort value by analyzing the differences between song characteristics and scene characteristics. For example, slow songs have a higher discomfort value in a bar scene.

[0229] In one possible implementation, the system calculates an inappropriateness value of 0.4 for slow songs and 0.1 for dance music in a family gathering scenario, thus optimizing the recommended content.

[0230] For example, in a bar setting, the discomfort value for slow songs is 0.5, while that for electronic music is 0.05, so the system reduces the proportion of slow songs recommended.

[0231] Step S84: Calculate the overall score of the song by using the time-weighted score and the scene inappropriateness value, and determine the sorting criteria.

[0232] Specifically, the overall score, combined with a time-weighted score and a scenario-appropriateness value, generates a final score for each song. The system then sorts the songs in descending order of their scores to generate a recommendation list.

[0233] In one embodiment, the system calculates a comprehensive score of 0.85 for dance music and 0.6 for popular songs, and generates a recommendation list that is primarily composed of dance music.

[0234] For example, in a bar setting, electronic music scores 0.9 overall, generating a recommendation list primarily featuring electronic music.

[0235] Step S85: Sort the songs in descending order according to their comprehensive scores, and use a quick sorting algorithm to generate a preliminary recommendation list.

[0236] Specifically, the quicksort algorithm sorts songs by their overall scores to generate a preliminary recommendation list.

[0237] In one possible implementation, the system sorts the 10 songs by their overall scores and generates a list that includes dance music and pop songs.

[0238] For example, in a bar setting, quick sorting generates a recommendation list primarily featuring electronic music, which improves user satisfaction.

[0239] Step S86: Update real-time preference data based on user feedback data, adjust the adaptive weight set, and optimize the recommendation list.

[0240] Specifically, user feedback data includes click-to-play and click-to-skip behaviors. The system updates the preference vector and weight set based on the feedback to optimize the recommendation list.

[0241] In one embodiment, the system reduces the weight of slow songs and optimizes the recommendation list based on user feedback that they skipped slow songs.

[0242] For example, in a bar setting, feedback shows that users prefer electronic music, so the system adjusts the weights and optimizes the recommended content.

[0243] Step S9: Obtain user preference features from historical playback records, generate user preference vectors through feature extraction algorithms, and obtain user preference representations.

[0244] Specifically, historical playback records reflect users' long-term preferences, and feature extraction algorithms generate preference vectors that represent users' inclinations towards song types.

[0245] In one possible implementation, the system analyzes user preferences for popular songs and generates a preference vector: popular songs 0.7, dance music 0.5.

[0246] For example, in a bar scenario, the user preference vector shows electronic music at 0.8, which improves recommendation accuracy.

[0247] Step S91: If the matching degree between the user preference representation and the song feature vector of the preliminary recommendation list is lower than a preset threshold, the user's recent playback behavior is analyzed by the short-term boredom pattern recognition algorithm to obtain the boredom pattern weight.

[0248] Specifically, a match score below a threshold, such as 0.7, indicates that the recommended list does not match the user's preferences. A short-term boredom pattern recognition algorithm analyzes recent behavior and calculates the weight of the boredom pattern.

[0249] In one embodiment, the system found that the user's match rate for popular songs was 0.6, which is lower than the threshold of 0.7, and the weight of the boredom pattern was 0.4.

[0250] For example, in a bar scenario, with a match rate of 0.65 and a boredom mode weight of 0.5, the system reduces the recommendation of slow songs.

[0251] Step S92: Using a reinforcement learning model, the song order in the initial recommendation list is adjusted based on the boredom mode weights and user preference representations to obtain the adjusted recommendation list.

[0252] Specifically, the reinforcement learning model optimizes the song order based on the boredom pattern weights, thereby improving the matching accuracy of the recommendation list.

[0253] In one possible implementation, the system adjusts the order of dance music, generates a new recommendation list, and improves the matching accuracy to 0.8.

[0254] For example, in a bar scenario, a reinforcement learning model can be used to sort electronic music tracks and optimize the recommendation performance.

[0255] Step S93: By comparing the matching degree between the adjusted recommendation list and the user preference representation using the matching degree calculation formula, determine whether it is higher than the preset threshold, and obtain the matching degree evaluation result.

[0256] Specifically, the matching degree calculation formula is based on cosine similarity, which compares the similarity between the recommendation list and the user preference vector.

[0257] In one embodiment, the adjusted recommendation list matching degree is 0.85, which is higher than the threshold of 0.7, indicating that the recommendation effect has improved.

[0258] For example, in a bar scenario, the match rate is 0.9, and the system confirms that the recommendation list is valid.

[0259] Step S94: If the matching degree evaluation result is lower than the preset threshold, the song feature weights are updated using the influence weight calculation formula to generate an updated recommendation list.

[0260] Specifically, if the matching degree is lower than the threshold, the system updates the song feature weights and regenerates the recommendation list.

[0261] In one possible implementation, the system updates the dance music weight to 0.8, generates a new recommendation list, and improves the matching accuracy.

[0262] For example, in a bar setting, the weight of electronic music is updated to 0.85 to optimize the recommendation list.

[0263] Step S95: Based on the updated recommendation list, the songs are reordered using a collaborative filtering algorithm to obtain an optimized recommendation sequence.

[0264] Specifically, the collaborative filtering algorithm reorders songs based on the updated weights to generate an optimized recommendation sequence.

[0265] In one embodiment, the system generates a recommendation sequence based on dance music, achieving a matching degree of 0.9.

[0266] For example, in a bar setting, optimizing the recommended sequence to primarily feature electronic music improves user satisfaction.

[0267] Step S96: By verifying the matching degree between the optimized recommended sequence and the historical playback records, determine whether the optimized recommended sequence meets the preset threshold, and obtain the final recommended list.

[0268] Specifically, the system verifies and optimizes the matching degree of the recommended sequences to ensure that the threshold requirements are met.

[0269] In one possible implementation, the optimized recommendation sequence matching degree is 0.92, which is higher than the threshold of 0.7, to generate the final recommendation list.

[0270] For example, in a bar setting, the final recommended list is dominated by electronic music, with a match rate of 0.95, which meets the user's needs.

[0271] Step S10: Obtain user interaction feedback data based on the recommendation sequence, and use time series analysis to extract the timestamps and interaction types of user behavior to obtain a user behavior dataset.

[0272] Specifically, interactive feedback data includes user behaviors such as clicking to play or skipping recommended sequences, and time series analysis is used to extract the temporal patterns of these behaviors.

[0273] In one embodiment, the system analyzes user feedback in a family gathering scenario to generate a behavior dataset that includes timestamps and interaction types.

[0274] For example, in a bar setting, user behavior datasets show that users frequently play electronic music after 10 p.m.

[0275] Step S101: Using the user behavior dataset, cluster analysis is used to divide the behavior categories, determine the classification boundaries between positive and negative behaviors, and obtain the behavior classification results.

[0276] Specifically, cluster analysis categorizes user behavior into positive behaviors, such as clicking to play, and negative behaviors, such as skipping.

[0277] In one possible implementation, the system categorizes behaviors into two types: play and skip, with the classification boundary based on behavior frequency.

[0278] For example, in a bar scenario, the classification results show that positive behaviors account for 70%, indicating that the recommendation effect is good.

[0279] Step S102: If the number of positive behaviors in the behavior classification results is greater than a preset threshold, then the weighted average method is used to fuse the behavior features to obtain the weight distribution of positive behaviors.

[0280] Specifically, if the number of positive behaviors exceeds a threshold, such as 50%, it indicates that the recommendation is effective. The system then integrates behavioral features to generate a weight distribution.

[0281] In one embodiment, positive behavior accounts for 60%, and the weight distribution shows that the weight of playback behavior is 0.7.

[0282] For example, in a bar scenario, the weight of positive behavior is 0.8, which optimizes the recommendation strategy.

[0283] Step S103: Based on the positive behavior weight distribution, extract the correlation of each behavior node in the logical chain, and use the Pearson correlation coefficient to analyze the correlation strength between nodes to obtain the behavior correlation matrix.

[0284] Specifically, logical chains represent causal relationships between behaviors, and Pearson correlation coefficients analyze the correlation between behavior nodes.

[0285] In one possible implementation, the system generates a behavior correlation matrix, showing that the correlation coefficient between playback and scene is 0.9.

[0286] For example, in a bar setting, the correlation matrix shows a strong correlation between electronic music playback and time period.

[0287] Step S104: Calculate the proportion of positive behaviors in the recommendation sequence using the behavior association matrix to determine the proportion of positive behaviors.

[0288] Specifically, the proportion of positive behavior reflects the effectiveness of the recommendation sequence.

[0289] In one embodiment, positive behaviors accounted for 75%, indicating that the recommended sequence was effective.

[0290] For example, in a bar scenario, positive behaviors account for 80%, and the system confirms that the recommendation is effective.

[0291] Step S105: If the proportion of positive behavior is lower than a preset threshold, adjust the parameters of the recommended sequence, reacquire user interaction feedback data, and obtain an updated behavior dataset.

[0292] Specifically, if the proportion of positive behaviors is below a threshold, such as 70%, the system adjusts the recommendation parameters and collects feedback again.

[0293] In one possible implementation, positive behaviors account for 65%, the system adjusts the weights, and generates a new behavior dataset.

[0294] For example, in a bar scenario, the system optimizes the recommendation sequence based on the 65% proportion of positive behavior.

[0295] Step S106: Based on the updated behavior dataset, repeat the behavior classification and proportion determination steps to obtain the final positive behavior proportion.

[0296] Specifically, the system repeatedly classifies and calculates proportions to ensure that the proportion of positive behaviors meets the requirements.

[0297] In one embodiment, the final positive behavior percentage was 78%, which met the threshold requirement.

[0298] For example, in a bar scenario, the final positive behavior rate was 82%, which optimized the recommendation effect.

[0299] Step S11: Calculate the proportion of positive behaviors by obtaining positive behavior records from the feedback data.

[0300] Specifically, positive behavior records include actions such as clicking to play, and the system calculates the proportion of such actions.

[0301] In one possible implementation, positive behavior accounts for 75%, providing a basis for model optimization.

[0302] For example, in a bar setting, the proportion of positive behaviors is 80%, indicating that the recommendation is effective.

[0303] Step S111: If the proportion of positive behavior exceeds a preset threshold, the sequence model parameters are initially adjusted to obtain the initial model parameters.

[0304] Specifically, if the proportion of positive behaviors exceeds a threshold, such as 70%, it indicates that the recommendation is effective, and the system adjusts the model parameters.

[0305] In one embodiment, the proportion of positive behavior is 78%, and the system adjusts the model weights to generate initial parameters.

[0306] For example, in a bar scene, the proportion of positive behaviors is 82%, and the system optimizes the model parameters.

[0307] Step S112: The initial model parameters are iteratively updated using an optimization algorithm to obtain the optimized model parameters.

[0308] Specifically, optimization algorithms, such as gradient descent, iteratively update the model parameters.

[0309] In one possible implementation, the system generates optimized model parameters through 10 iterations.

[0310] For example, in a bar scenario, optimizing parameters improves the model's prediction accuracy.

[0311] Step S113: Update the sequence model by optimizing the model parameters to generate a behavior change prediction model.

[0312] Specifically, optimizing model parameters is used to update the sequence model and generate a new prediction model.

[0313] In one embodiment, the new model improves the accuracy of predicting user preference for dance music to 90%.

[0314] For example, in a bar setting, the new model predicted electronic music preferences with 92% accuracy.

[0315] Step S114: Obtain the behavioral change trend for the next period based on the behavioral change prediction model.

[0316] Specifically, behavior change prediction models predict users' future behavioral trends.

[0317] In one possible implementation, the model predicts the user's preferred dance music for the following week.

[0318] For example, in a bar setting, the model predicts an increasing trend of users preferring electronic music.

[0319] Step S115: If the trend of behavior change meets the preset conditions, the trend data is input into the cycle prediction module to obtain the prediction result of the next cycle.

[0320] Specifically, the preset condition can be that the trend strength is higher than 0.8, and the system generates cycle prediction results.

[0321] In one embodiment, the trend strength is 0.85, and the prediction results show that users prefer dance music.

[0322] For example, in a bar setting, the predictions show that users will continue to prefer electronic music.

[0323] Step S116: Adjust the feedback data processing strategy based on the periodic prediction results to generate an updated data processing flow.

[0324] Specifically, the periodic prediction results optimize the data processing flow and improve recommendation efficiency.

[0325] In one possible implementation, the system adjusts the feedback processing frequency to optimize the recommendation effect.

[0326] For example, in a bar setting, the updated process improved the accuracy of electronic music recommendations.

[0327] In the description of this invention, it should be understood that the terms "coaxial," "bottom," "one end," "top," "middle," "other end," "upper," "side," "top," "inner," "front," "center," "both ends," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0328] Furthermore, the terms “first,” “second,” “third,” and “fourth” are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as “first,” “second,” “third,” or “fourth” may explicitly or implicitly include at least one of those features.

[0329] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "setting," "connection," "fixing," "screw connection," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal connection of two components or the interaction between two components. Unless otherwise explicitly limited, those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0330] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A response optimization method based on a karaoke machine, characterized in that, include: Based on the collected user click play and skip behavior data on the karaoke machine, combined with time period and activity scenario information, behavior sequence features are obtained. After processing the behavior sequence features using behavior sequence temporal coding, dynamic change indicators of user behavior are obtained. Based on the aforementioned dynamic change indicators of user behavior, a sequence model is used to analyze the time-related patterns in the behavior sequence. Through the time pattern extraction mechanism and dynamic indicator quantification, the trend of behavior change is obtained. If the behavioral change trend shows that the skip rate for a certain type of song is higher than a preset threshold, then the corresponding time period and scene factors are extracted, and the skip rate threshold comparison and context parsing rule definition are used to determine whether the change is due to short-term boredom or scene discomfort to obtain the context influence weight. Based on the context influence weights and the rate of change of trend prediction error correction behavior, the song type preference coefficients in the fixed weights are evaluated and updated, and an adaptive weight set is obtained through pattern correlation analysis and index fusion weight adjustment. The comprehensive score of the songs is calculated using the adaptive weight set, and the ranking is optimized based on the current user's real-time preferences. The analysis of time period factors and the quantification of scene incompatibility indicators are incorporated to obtain a preliminary recommendation list. If the matching degree between the songs in the preliminary recommendation list and the historical playback records is lower than a preset threshold, a reinforcement learning model is used to fine-tune the order of the songs in the preliminary recommendation list, and the short-term boredom pattern recognition and influence weight calculation formula are used to determine whether the adjusted list improves the matching degree, so as to obtain an optimized recommendation sequence. Based on the optimized recommendation sequence, user subsequent interaction feedback data is obtained and combined with the factor extraction logic chain and judgment basis data fusion to obtain the proportion of positive behavior in the feedback data. Based on the proportion of positive behavior in the feedback data, the parameters of the sequence model are updated, and after optimization of the sequence model parameters, a behavior change prediction model for the next cycle is obtained.

2. The response optimization method based on a karaoke machine according to claim 1, characterized in that: The step of obtaining dynamic change indicators of user behavior by collecting user click-play and skip behavior data on the karaoke machine, combining time period and activity scenario information to obtain behavioral sequence features, and then processing the behavioral sequence features using behavioral sequence temporal coding includes: By collecting user click play and skip behavior data on the karaoke machine, and combining it with time period and activity scenario, user behavior sequences are generated; The behavior sequence is processed using a time window segmentation method to extract sequence features and obtain the temporal distribution features of the behavior sequence; If the frequency of clicks to play in the time distribution features is higher than a preset threshold, then a long short-term memory network is used to perform temporal encoding on the behavior sequence to obtain the encoded behavior sequence vector. Based on the encoded behavior sequence vector, the similarity of user behavior patterns in different time periods and activity scenarios is calculated to obtain behavior pattern clustering results. Based on the clustering results of the aforementioned behavioral patterns, the dynamic change trends of users in specific activity scenarios are extracted to obtain dynamic change indicators. Based on the aforementioned dynamic change indicators, combined with time period and scenario associations, a user behavior preference sequence is generated to obtain a user behavior prediction model.

3. The response optimization method based on a karaoke machine according to claim 1, characterized in that: The step of analyzing time-related patterns in a behavior sequence using a sequence model based on the user behavior dynamic change indicators, and obtaining the behavior change trend through a time pattern extraction mechanism and dynamic indicator quantification calculation includes: User behavior sequence data is acquired, and time series decomposition method is used to extract time-related patterns and obtain the time features of the behavior sequence. By using time series features and employing a long short-term memory network model, we can analyze the dynamic change patterns in behavioral sequences and obtain behavioral pattern prediction results. If the behavior pattern prediction results deviate significantly from the preset threshold, the sliding window method is used to calculate the dynamic index quantification value to obtain the quantitative characteristics of behavior changes. Based on the quantitative characteristics of the behavioral changes, a decision tree model is used to determine the category of the behavioral change trend and obtain the trend classification result. Based on the trend classification results and combined with the time-related pattern, a weighted average method is used to calculate the intensity of the behavioral change trend and obtain the trend intensity value. The trend strength value is judged. If the strength value exceeds a preset threshold, the stability of the behavior change trend is obtained through clustering method, thereby obtaining trend stability characteristics. Based on the stability characteristics of the trend, a time series smoothing method is used to optimize the output of the behavioral change trend and obtain the final behavioral change trend.

4. The response optimization method based on a karaoke machine according to claim 1, characterized in that: If the behavioral change trend shows that the skip rate for a certain type of song is higher than a preset threshold, the step of extracting the corresponding time period and scenario factors, and using skip rate threshold comparison and context parsing rules to determine whether the change stems from short-term boredom or scenario discomfort to obtain the context influence weight includes: Based on the skip rate data of various songs obtained by users, statistical analysis is used to calculate the mean and variance of the skip rate and obtain the trend of skip rate changes. If the skip rate trend is higher than the preset threshold, the context parsing module will extract the scene factors and time period factors during user interaction to generate a context feature set. The context feature set is classified using a decision tree algorithm to determine whether the change in skip rate stems from short-term boredom or scene discomfort, and the classification result is obtained. Based on the classification results, the contextual impact weights of short-term boredom and scene discomfort are calculated, and a weighted average method is used to generate a comprehensive impact weight. The recommendation priority of song categories is adjusted by the comprehensive influence weights to generate an updated recommendation sequence; Based on the updated recommendation sequence, the A / B testing module is used to verify the recommendation effect, obtain user interaction data and the optimized skip rate; If the optimized skip rate is still higher than the preset threshold, the interaction data is input into the context parsing module through the feedback loop module to regenerate the context feature set.

5. The response optimization method based on a karaoke machine according to claim 1, characterized in that: The step of updating the song type preference coefficient in the fixed weights based on the context influence weights and the rate of change of trend prediction error correction behavior, and obtaining an adaptive weight set through pattern correlation analysis and index fusion weight adjustment, includes: Historical data is obtained from user interaction records, and user preferences are extracted using time series analysis to obtain a set of user preference features; Based on the user preference feature set, the K-means clustering algorithm is used to divide the song types and obtain song type groups; By grouping by song type, the rate of behavioral change is calculated to obtain the trend of behavioral change. Based on the trend of behavioral change, if the rate of change exceeds a preset threshold, a linear regression algorithm is used to predict the trend deviation and obtain the trend prediction error. By analyzing the trend prediction error and combining it with pattern correlation analysis, the fixed weights are adjusted to obtain the weight update coefficients. Based on the weight update coefficients and the fusion index fusion method, the adaptive weight set is updated to obtain the optimized weight set; By optimizing the weight set and reallocating the recommendation priority of song types, the final recommendation ranking is obtained.

6. The response optimization method based on a karaoke machine according to claim 1, characterized in that, The steps of calculating the comprehensive score of songs using the adaptive weight set, optimizing the ranking based on the current user's real-time preferences, incorporating time-time factor decomposition and scene incompatibility index quantification, and obtaining a preliminary recommendation list include: Real-time preference data is obtained based on users' historical interaction data and current behavior, and user preference vectors are obtained through collaborative filtering algorithms. Based on the real-time preference vector and song metadata, an adaptive weight set is calculated to obtain the weight allocation for each song; By analyzing the characteristics of the current time period through a time period segmentation model, and combining this with user preference vectors, the weight set is adjusted to generate a time-weighted score. A scene discomfort index quantitative model is used to evaluate the matching degree between the song and the current scene, and the scene discomfort value is calculated. The overall score of a song is calculated by combining time-weighted scores and scene-inappropriateness values, which forms the basis for ranking. Songs are sorted in descending order based on their overall scores, and a preliminary recommendation list is generated using a quick sorting algorithm. We update real-time preference data based on user feedback, adjust the adaptive weight set, and optimize the recommendation list.

7. The response optimization method based on a karaoke machine according to claim 1, characterized in that: The step of obtaining the proportion of positive behavior in the feedback data by combining the user's subsequent interaction feedback data obtained from the optimized recommendation sequence with the factor extraction logic chain and judgment basis data fusion includes: We acquire user interaction feedback data based on recommendation sequences, and use time series analysis to extract timestamps and interaction types of user behavior to obtain a user behavior dataset. Based on the user behavior dataset, cluster analysis is used to divide the behavior into categories, determine the classification boundaries between positive and negative behaviors, and obtain the behavior classification results. Based on the positive behavior weight distribution, the correlation of each behavior node in the logical chain is extracted, and the correlation strength between nodes is analyzed by Pearson correlation coefficient to obtain the behavior correlation matrix; The proportion of positive behaviors in the recommendation sequence is calculated by using the behavior association matrix. If the proportion of positive behavior is lower than the preset threshold, the parameters of the recommended sequence are adjusted, user interaction feedback data is reacquired, and an updated behavior dataset is obtained. Based on the updated behavior dataset, repeat the behavior classification and proportion determination steps to obtain the final positive behavior proportion.

8. A response optimization system based on a karaoke machine, characterized in that, include: The behavior sequence feature acquisition module is configured to acquire behavior sequence features based on the collected user click play and skip behavior data on the karaoke machine, combined with time period and activity scenario information, and then obtain the user behavior dynamic change index by processing the behavior sequence features with behavior sequence temporal coding. The user behavior dynamic change index generation module is configured to analyze the time-related patterns in the behavior sequence using a sequence model based on the user behavior dynamic change index, and obtain the behavior change trend through a time pattern extraction mechanism and dynamic index quantification calculation. The behavior change trend determination module is configured to extract the corresponding time period and scene factors if the behavior change trend shows that the skip rate of a certain type of song is higher than a preset threshold, and use the skip rate threshold comparison and context parsing rule definition to determine whether the change is due to short-term boredom or scene discomfort to obtain the context influence weight. The context influence weight acquisition module is configured to update the song type preference coefficient in the fixed weight based on the context influence weight and the rate of change of trend prediction error correction behavior, and to obtain an adaptive weight set through pattern correlation analysis and index fusion weight adjustment. The adaptive weight set generation module is configured to calculate the comprehensive score of the songs through the adaptive weight set, optimize the sorting based on the current user's real-time preferences, incorporate time period factor decomposition and scene incompatibility index quantification, and obtain a preliminary recommendation list. The initial recommendation list determination module is configured to, if the matching degree between the songs in the initial recommendation list and the historical playback records is lower than a preset threshold, use a reinforcement learning model to fine-tune the order of the songs in the initial recommendation list, and use short-term boredom pattern recognition and influence weight calculation formula to determine whether the adjusted list improves the matching degree, thereby obtaining an optimized recommendation sequence. The optimized recommendation sequence acquisition module is configured to obtain the proportion of positive behavior in the feedback data by combining the user's subsequent interaction feedback data with the factor extraction logic chain and judgment basis data fusion based on the optimized recommendation sequence. The behavior change prediction model update module is configured to update the parameters of the sequence model based on the proportion of positive behavior in the feedback data, and incorporate the optimized sequence model parameters to obtain the behavior change prediction model for the next period.

9. A karaoke device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the response optimization method based on a karaoke machine as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the response optimization method based on a karaoke machine as described in any one of claims 1 to 7.

Images