Intelligent music recommendation method based on emotion perception and acoustic characteristics

Abstract

The invention discloses an intelligent music recommendation method based on emotional perception and acoustic features, and relates to the technical field of intelligent recommendation systems and emotional computation.The method comprises the steps that physiological signals, music acoustic features and historical behavior data of a user are collected; preprocessing the multi-source features and mapping the multi-source features to the same dimension to construct a fusion matrix; building a double-branch deep learning model, fusing features through an attention mechanism and training parameters; an evolutionary algorithm is adopted to optimize hyper-parameter screening optimal combination; generating a recommendation list matched with the real-time emotion and preference; and continuously collecting user interaction data, and regularly and incrementally training the dynamic update model. According to the method, emotion and behavior dual-drive recommendation is achieved by fusing physiological signals and acoustic features, emotion perception is accurate, recommended content fits the real-time mood, model optimization is efficient, recommendation precision and diversity are remarkably improved, and the music consumption experience of a user is greatly improved.

Description

Technical Field

[0001] This invention relates to the fields of intelligent recommendation systems and emotion computing technology, and in particular to an intelligent music recommendation method based on emotion perception and acoustic features. Background Technology

[0002] With the rapid development of artificial intelligence and digital technology, the cultural consumption market is experiencing product overload. Music, as a core cultural carrier of emotional expression, has undergone a fundamental transformation in its dissemination and consumption patterns. Users are no longer satisfied with passively receiving music content; instead, they crave personalized recommendations that closely match their real-time mood. The value of music is increasingly intertwined with the user's psychological state, making intelligent recommendation technology with emotional perception capabilities a key breakthrough for improving user experience. Currently, mainstream music recommendation systems primarily rely on two technical paradigms: collaborative filtering and deep learning. Collaborative filtering generates recommendations by mining similarities between user groups or musical works, but its core limitation lies in its over-reliance on historical behavioral data, making it difficult to effectively integrate multi-source heterogeneous information such as real-time physiological signals and textual emotional expressions, resulting in recommendations lagging behind changes in user emotions.

[0003] While deep learning models possess the potential for multimodal data processing, industrial practice still heavily relies on structured behavioral data such as playback history and collection preferences, failing to adequately mine the unstructured emotional features inherent in music itself. Musical emotional expression is often conveyed through acoustic features such as tempo, pitch, loudness, and timbre spectrum. Current technologies have not fully extracted these core features and established a correlation mapping with user emotions. Furthermore, the lack of a systematic incorporation of user physiological and psychological characteristics prevents recommendation systems from accurately capturing users' deep emotional needs. A user's emotional state is directly reflected through physiological signals such as heart rate, skin conductivity, and respiratory rate. However, existing recommendation systems generally ignore this crucial information, relying solely on historical behavior to infer preferences. This leads to frequent misalignments between recommendation results and the user's current mood, making it difficult to achieve "emotionally resonant" recommendations.

[0004] Furthermore, existing recommendation systems primarily rely on grid search methods for hyperparameter optimization. This method requires traversing all possible parameter combinations, resulting in high computational complexity, low optimization efficiency, and a tendency to get trapped in local optima, further limiting the performance improvement of recommendation models. These combined technical bottlenecks lead to insufficient accuracy and a lack of personalization in existing recommendation systems, failing to accurately respond to users' dynamically changing emotional needs. This not only reduces the user's music consumption experience but also hinders the effective dissemination of music content. Given the growing demand for personalized and emotional services, traditional "behavior-driven" recommendation models are no longer sufficient to meet market demands. There is an urgent need to develop an intelligent recommendation technology that integrates emotion perception and acoustic feature analysis, driving music recommendation from "behavior-driven" to "emotion + behavior-driven," achieving a dual improvement in recommendation accuracy and user experience. Summary of the Invention

[0005] This invention proposes an intelligent music recommendation method based on emotion perception and acoustic features to solve the problems mentioned in the prior art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: An intelligent music recommendation method based on emotion perception and acoustic features includes the following steps: Multi-source data is collected, including real-time collection of physiological signals such as heart rate, skin conductivity, respiratory rate, and body temperature via wearable sensing devices; acoustic features of the music to be recommended are extracted using audio analysis tools, including tempo, pitch, loudness, timbre spectrum, and melody outline; and structured user behavior data is acquired simultaneously. Preprocessing of multi-source features involves denoising, smoothing, and normalizing physiological signals to remove abnormal fluctuations; standardization transformation of acoustic features; encoding of structured behavioral data; and mapping of the three types of features to the same dimensional space using feature alignment techniques to construct a fusion feature matrix. A dual-branch deep learning model is constructed, consisting of an emotion perception branch and an acoustic feature branch. The emotion perception branch uses a convolutional neural network to extract emotion-related features from physiological signals, while the acoustic feature branch uses a recurrent neural network to capture the temporal correlation of musical acoustic features. The output features of the two branches are fused through an attention mechanism, and the model parameters are trained by combining user structured behavior data. Evolutionary algorithms are used to optimize model hyperparameters. The hyperparameters to be optimized include learning rate, number of network layers, number of neurons, and attention weight coefficient. The hyperparameter population is initialized. The model recommendation accuracy is used as the fitness function. The population is iteratively updated through selection, crossover, and mutation operations to select the optimal combination of hyperparameters and configure it into the model. A real-time recommendation list is generated. The model receives the user's real-time physiological signals and the acoustic features of the currently playing music, and outputs the user's real-time emotion assessment results and music suitability scores. Based on the assessment results and scores, a recommendation list containing 10-20 songs is generated. The model is dynamically updated, and real-time user feedback behavior is collected, including playback duration, favorites, sharing, and skipping. The model is incrementally trained regularly by combining new physiological signals and acoustic feature data to optimize feature extraction weights and model parameters.

[0007] Furthermore, it also includes an emotion intensity quantification step, which calculates the user's real-time emotion intensity by combining the time-domain and frequency-domain characteristics of various physiological signals. The calculation expression is as follows: ;in This is a quantitative value for emotional intensity, ranging from 0 to 10. For heart rate feature weights, This is the normalized value of the heart rate coefficient of variation. Weighted by skin conductivity, The result is a normalized value for fluctuations in skin conductivity. As a weight for respiratory rate, This is the normalized value of the rate of change in respiratory rate. Weighted by body temperature The normalized result of body temperature deviating from the baseline value; .

[0008] Furthermore, it also includes a dynamic allocation step for acoustic feature weights, adjusting the contribution weights of each acoustic feature based on the user's emotional state and music style preferences, calculated as follows: ;in For the first The final weights of the acoustic features; The emotion correlation coefficient ranges from 0.1 to 0.5. For the first The correlation between acoustic features and current mood; For the first The fundamental importance weights of acoustic features; This is a summation operation for all acoustic features.

[0009] Furthermore, it also includes model feature extraction optimization steps, introducing an attention gating mechanism in the emotion perception branch to assign differentiated attention weights to the features of physiological signals at different time periods, focusing on extracting signal features during periods of intense emotional fluctuation; adding a spectrogram attention module in the acoustic feature branch to focus on acoustic frequency band features that are highly correlated with user emotions; and enhancing the model's ability to capture key features through a dual attention mechanism to reduce interference from irrelevant features.

[0010] Furthermore, it also includes a real-time emotion dynamic tracking step, which sets a sliding time window to continuously collect physiological signals, with the window duration set to 30 seconds to 5 minutes and dynamically adjustable; calculates the emotion intensity quantification value every 5 seconds to generate an emotion change curve; when the emotion change amplitude exceeds a preset threshold, it triggers the model to update the recommendation strategy in real time, adjusts the acoustic feature adaptation direction, and realizes that the recommendation results are adjusted in real time to follow the user's emotion changes.

[0011] Furthermore, it includes steps to enhance user personalization preferences, such as constructing a user preference vector library and mapping user historical behavior data, long-term emotional characteristics, and acoustic feature preferences into high-dimensional vectors; calculating the cosine similarity between newly recommended music and preference vectors, and generating a comprehensive recommendation index by combining emotional fit scores; music with a comprehensive recommendation index higher than a set threshold is given priority in the recommendation list; and user emotional feedback on different styles of music is recorded to enhance the extraction and application of personalized preference features.

[0012] Furthermore, it also includes a multimodal feedback fusion step, which collects multimodal feedback information from the user terminal, including semantic evaluation text, facial expression images, and operational behaviors; performs sentiment analysis on the semantic evaluation text to extract text emotion features; performs facial feature recognition on the facial expression images to obtain visual emotion features; and fuses the text emotion features, visual emotion features, and physiological signal emotion features to correct the model's emotion evaluation results and recommendation strategy, forming a multi-dimensional feedback closed loop.

[0013] Furthermore, it also includes iterative optimization steps for the evolutionary algorithm. The selection operation for the hyperparameter population is optimized using the roulette wheel selection method; the crossover operation uses a single-point crossover strategy; the mutation operation uses an adaptive mutation probability, which is dynamically adjusted with the number of iterations. In the early stage of the iteration, a higher mutation probability is set to explore the hyperparameter space, and in the later stage of the iteration, the mutation probability is reduced to focus on the optimal solution. After each iteration, the fitness value of each individual in the population is calculated, and the top 30% of individuals with the highest fitness are retained to enter the next generation of the population. The iteration stops when the preset number of iterations or the fitness value converges, and the optimal hyperparameter combination is output.

[0014] Furthermore, it also includes a music style segmentation and adaptation step, which subdivides music into different subcategories according to style; constructs a unique acoustic feature template for each style subcategory; calculates the compatibility probability between the current emotion and each style subcategory by combining the user's emotional intensity and style preference; prioritizes recommending music from the style subcategory with the highest compatibility probability; and sorts the music by acoustic feature matching degree within the same style subcategory.

[0015] Furthermore, it also includes a diversity balancing step, which calculates the acoustic feature similarity of different music in the recommendation list; when the average similarity is higher than a set threshold, a diversity adjustment factor is introduced to adjust the ranking results and reduce the ranking priority of highly similar music; and tracks with greater acoustic feature differences are selected from music with high suitability to supplement the recommendation list.

[0016] Compared with existing technologies, the beneficial effects of this invention are: The intelligent music recommendation method based on emotion perception and acoustic features of this invention achieves a comprehensive innovation to address the technical pain points of existing recommendation systems. Through multi-source data fusion, model structure optimization, and intelligent algorithm innovation, it significantly improves the emotion perception capability, feature extraction accuracy, and personalized adaptation level of the recommendation system, providing users with music recommendation services that better match their mood.

[0017] In terms of innovative recommendation logic, the method breaks through the limitations of traditional "behavior-driven" approaches, constructing a dual-core recommendation framework of "emotion + behavior-driven." This framework deeply integrates user physiological signals, musical acoustic features, and historical behavioral data, achieving a leap from "predicting user preferences" to "perceiving user emotions." By accurately collecting physiological signals such as heart rate and skin conductivity, and extracting acoustic features such as rhythm and pitch, a mapping relationship between emotions and musical features is established. This allows recommendation results to move beyond simply continuing historical behavior and instead match the user's current mood in real time, significantly improving the relevance and emotional resonance of recommended content.

[0018] In terms of model performance optimization, the dual-branch deep learning model enhances the extraction capabilities of physiological emotional features and musical acoustic features, respectively. The multi-head attention mechanism effectively focuses on key features, reduces interference from irrelevant information, and significantly improves the targeting and effectiveness of feature extraction. The introduction of evolutionary algorithms replaces the traditional grid search method, reducing the complexity and computational cost of hyperparameter optimization, quickly selecting the optimal hyperparameter combination, and ensuring the model maintains high efficiency, thus improving the accuracy and response speed of the recommendation system from a technical perspective. Simultaneously, the multimodal feedback closed-loop mechanism integrates feedback information such as text evaluation, facial expressions, and user actions, continuously refining the model's emotion assessment and recommendation strategy, driving dynamic improvement in recommendation accuracy.

[0019] In terms of user experience enhancement, the real-time emotion tracking function can keenly capture changes in user emotions and adjust recommendation directions in a timely manner, ensuring that recommendation results are synchronized with emotional fluctuations. The user personalization preference enhancement step deeply mines individual style preferences and emotional feedback patterns, ensuring that recommended content not only matches the current mood but also aligns with long-term aesthetics, achieving an organic unity of personalization and emotional engagement. The recommendation diversity balancing mechanism effectively avoids aesthetic fatigue caused by content monotony, enriching recommendation dimensions while ensuring emotional adaptability and improving the long-term user experience. Overall, the method comprehensively addresses the problems of insufficient emotion perception, weak feature extraction, and difficulties in hyperparameter optimization in existing recommendation systems through multi-dimensional technological innovation, making music recommendations more intelligent, accurate, and emotionally resonant, satisfying users' deep emotional needs and providing technical support for the precise dissemination of music content. Attached Figure Description

[0020] Figure 1 This is a schematic block diagram of an intelligent music recommendation method based on emotion perception and acoustic features proposed in this invention. Figure 2 A comparison chart of recommendation matching accuracy under different emotion types; Figure 3 A graph showing the percentage contribution of different features to recommendation accuracy; Figure 4 Performance comparison chart of the model before and after hyperparameter optimization; Figure 5 User satisfaction evaluation charts for different application scenarios; Figure 6 This is a chart comparing the proportions of recommended content. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," 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.

[0023] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. Furthermore, the terms "installed," "connected," and "linked" should be interpreted broadly; for example, they may refer to a fixed connection, a detachable connection, or an integral connection; they may refer to a mechanical connection or an electrical connection; they may refer to a direct connection or an indirect connection through an intermediate medium; and they may refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. The invention will now be described in further detail with reference to the accompanying drawings.

[0024] Reference Figures 1 to 6 A smart music recommendation method based on emotion perception and acoustic features includes the following steps: Intelligent music recommendation methods based on emotion perception and acoustic features include: Collect multi-source data, including real-time collection of physiological signals such as heart rate, skin conductivity, respiratory rate, and body temperature through wearable sensing devices, and extract acoustic features of recommended music such as tempo, pitch, loudness, timbre spectrum, and melody outline through professional audio analysis tools. Simultaneously acquire structured behavioral data such as users' historical music playback records, favorites preferences, skipping behavior, and playback completion rate. Multi-source features are preprocessed, and physiological signals are denoised by moving average, smoothed by Gaussian filtering, and normalized by min-max. Abnormal fluctuation data exceeding 3 times the standard deviation are removed. Acoustic features are transformed by Z-score standardization. Structured behavioral data are encoded one-hot. The three types of features are mapped to a 512-dimensional unified space through feature alignment technology to construct a fusion feature matrix. A dual-branch deep learning model is constructed, with an emotion perception branch and an acoustic feature branch built separately. The emotion perception branch uses a one-dimensional convolutional neural network to extract emotion-related features from physiological signals, while the acoustic feature branch uses a long short-term memory network to capture the temporal correlation of musical acoustic features. The output features of the two branches are fused through a multi-head attention mechanism, and the model parameters are trained using a cross-entropy loss function in combination with user structured behavior data. Evolutionary algorithms were used to optimize the model hyperparameters. Genetic algorithm was selected as the core evolutionary algorithm. The hyperparameters to be optimized included learning rate, number of network layers, number of neurons, and attention weight coefficient. The initial size of the hyperparameter population was 50. The fitness function was the model recommendation accuracy. The population was updated iteratively through selection, single-point crossover, and adaptive mutation operations. The number of iterations was set to 100 generations. The optimal combination of hyperparameters was selected and configured into the model. The model generates a real-time recommendation list. It receives the user's real-time physiological signals and the acoustic features of the current music playback, outputs the user's current emotional state assessment results and music suitability score, and combines the user's historical preferences and real-time scenario requirements to generate a recommendation list containing 15 songs in descending order of suitability score. The list is then synchronously fed back to the user's terminal and the reasons for the suitability are displayed. The model is dynamically updated, continuously collecting user interaction behaviors related to recommendation results, including playback duration, favorites, sharing, skipping, and repeat playback. Combined with new physiological signals and acoustic feature data, the model is incrementally trained every 24 hours to optimize feature extraction weights and model parameters, dynamically adapting to changes in user emotions and preference shifts.

[0025] This invention also includes an emotion intensity quantification step, which calculates the user's real-time emotion intensity by combining the time-domain and frequency-domain characteristics of various physiological signals. The calculation expression is as follows: in The emotional intensity quantification value ranges from 0 to 10. The weights for the heart rate feature range from 0.2 to 0.3. The normalized value of the heart rate coefficient of variation ranges from 0 to 1. The weighting value for skin conductivity ranges from 0.25 to 0.35. The normalized values ​​for skin conductivity fluctuations range from 0 to 1. The weighting for respiratory rate ranges from 0.2 to 0.3. The normalized value of the rate of change of respiratory rate ranges from 0 to 1. The weighting value for body temperature ranges from 0.15 to 0.25. The normalized result for body temperature deviation from the baseline value ranges from 0 to 1. This calculation enables the deep fusion of emotional information from multiple physiological signals, accurately quantifying the intensity of a user's current emotions.

[0026] This invention also includes a dynamic allocation step for acoustic feature weights, adjusting the contribution weights of each acoustic feature based on the user's emotional state and music style preference, with the calculation expression being: ,in For the first The final weights of the acoustic features. The emotion correlation coefficient ranges from 0.1 to 0.5. For the first The correlation between acoustic features and current emotion ranges from 0 to 1. For the first The basic importance weights for acoustic features range from 0.1 to 0.3. The summation operation of all acoustic features allows the acoustic feature weights to be dynamically adjusted according to the user's real-time emotions, making the music feature matching more in line with the user's current mood.

[0027] This invention also includes a model feature extraction optimization step. A spatial attention gating mechanism is introduced into the emotion perception branch to assign differentiated attention weights to the features of physiological signals at different time periods, focusing on extracting signal features during periods of intense emotional fluctuation. In the acoustic feature branch, an attention module based on Mel spectrograms is added to focus on acoustic frequency band features that are highly correlated with user emotions. The dual attention mechanism enhances the model's ability to capture key features, weakens the interference of irrelevant features on recommendation results, and improves the targeting and effectiveness of feature extraction.

[0028] This invention also includes a real-time emotion dynamic tracking step, which sets a sliding time window to continuously collect physiological signals. The default window duration is set to 2 minutes and allows users to adjust it independently within the range of 30 seconds to 5 minutes. The emotion intensity quantification value is calculated every 5 seconds, and an emotion change curve is generated and monitored in real time. When the emotion change amplitude exceeds 2 in two consecutive calculation cycles, the model is triggered to update the recommendation strategy in real time, adjusting the acoustic feature adaptation direction and weight allocation ratio to ensure that the recommendation results are adjusted in real time to follow the user's emotion changes.

[0029] This invention also includes a user personalized preference enhancement step, which involves constructing a user preference vector library, mapping user historical behavior data, long-term emotional features, and acoustic feature preferences into a 1024-dimensional high-dimensional vector, calculating the matching degree between new recommended music and the preference vector using normalized cosine similarity, and generating a comprehensive recommendation index by combining the emotional fit score with a weight ratio of 0.6:0.4. Music with a comprehensive recommendation index higher than 8.0 is given priority in the recommendation list. At the same time, user emotional feedback on different styles of music is recorded to continuously strengthen the extraction and application of personalized preference features and deepen the personalization of recommendations.

[0030] This invention also includes a multimodal feedback fusion step, which collects multimodal feedback information from user terminals, including semantic evaluation text, facial expression images, and operational behaviors. It uses BERT-based semantic evaluation text sentiment analysis to extract text emotion features, and uses facial key point detection technology to perform facial feature recognition on facial expression images to obtain visual emotion features. It then fuses text emotion features, visual emotion features, and physiological signal emotion features with equal weights to correct the model's emotion evaluation results and recommendation strategy, forming a multi-dimensional feedback closed loop to continuously optimize recommendation accuracy.

[0031] This invention also includes an evolutionary algorithm iterative optimization step. The selection operation of the hyperparameter population is optimized using a roulette wheel selection method that allocates selection probabilities according to the proportion of fitness values. The crossover operation adopts a single-point crossover strategy with a crossover probability set to 0.7. The mutation operation adopts an adaptive mutation probability, which is dynamically adjusted with the number of iterations. In the early stage of the iteration, a relatively high mutation probability of 0.3 is set to explore the hyperparameter space, and in the later stage of the iteration, it is gradually reduced to 0.05 to focus on the optimal solution. After each iteration, the fitness value of each individual in the population is calculated. The individuals with the top 30% fitness values ​​are directly entered into the next generation population and new individuals are added. The iteration stops when the fitness value fluctuation is less than 0.001 for a preset number of iterations or 10 consecutive generations, and the optimal hyperparameter combination is output.

[0032] This invention also includes a music style segmentation and adaptation step, which subdivides music into subcategories such as classical, pop, rock, jazz, electronic, folk, and R&B. For each style subcategory, a unique acoustic feature template is constructed. The classical music template focuses on low-frequency timbre spectrum and slow tempo, while the electronic music template focuses on high-frequency loudness and fast tempo. Combining the user's emotional intensity and style preference, a logistic regression algorithm is used to calculate the compatibility probability between the current emotion and each style subcategory. Music from the style subcategory with the highest compatibility probability is recommended first. At the same time, within the same style subcategory, music is sorted by acoustic feature matching degree to ensure that the recommendation results meet both emotional needs and style preferences.

[0033] This invention also includes a diversity balancing step, which uses Euclidean distance to calculate the acoustic feature similarity of different music in the recommendation list. When the average similarity is higher than 0.7, the ranking results are adjusted by a diversity adjustment factor. The diversity adjustment factor ranges from 0.1 to 0.3. The ranking priority of highly similar music is reduced according to the adjustment factor weight. Tracks with significant differences in acoustic features are selected from the top 30 music in terms of suitability score to supplement the recommendation list. While ensuring emotional suitability and personalization, the diversity of the recommendation list is improved, avoiding user aesthetic fatigue caused by monotonous recommendation content.

[0034] The present invention will be further illustrated below through two embodiments: Example 1 Smart music recommendation application for everyday mobile devices This embodiment is applied to mobile terminals such as mobile phones and tablets, and is suitable for diverse scenarios such as daily commuting, home leisure, and office breaks. It is aimed at ordinary music consumers and combines wearable sensing devices to achieve accurate emotion perception and acoustic feature matching. It fully implements all technical solutions and creates a personalized music recommendation experience driven by "emotion and behavior".

[0035] During the multi-source data collection phase, physiological signals are collected in real time through wearable devices worn by the user. The sampling frequency is set to 10Hz, and four core indicators—heart rate, skin conductivity, respiratory rate, and body temperature—are continuously collected, generating a set of average data every 30 seconds. The device's built-in professional audio analysis tool extracts acoustic features from 5000 songs in the recommended music library, including tempo (60-180 BPM), pitch (C1-B7), loudness (40-90 dB), timbre spectrum (20-20000 Hz), and melody outlines categorized by pitch variation trends. Simultaneously, the user's historical music data for the past six months is acquired, including playback history, favorites preferences, skipping behavior, and playback completion rate. Playback records exceeding 30 seconds without skipping are marked as valid.

[0036] In the preprocessing stage of multi-source features, the physiological signals are denoised using a moving average method with a window size of 5. Gaussian filtering is used to smooth data fluctuations, and min-max normalization is applied to map the data to the 0-1 interval, removing outliers exceeding three standard deviations. Acoustic features are Z-score standardized to eliminate dimensional differences. Structured behavioral data are one-hot encoded, with 1 marked as "favorited" and 0 as "unfavorited." Feature alignment technology is used to uniformly map the three types of features to a 512-dimensional space, constructing an N×512 fusion feature matrix, where N is the number of music tracks to be recommended.

[0037] In the dual-branch deep learning model construction phase, the emotion perception branch employs a 3-layer one-dimensional convolutional neural network with kernel sizes of 3, 5, and 7, using ReLU as the activation function to extract emotion-related features from physiological signals. The acoustic feature branch uses a 2-layer long short-term memory network with 256 hidden units per layer to capture the temporal correlation of acoustic features. The dual-branch output features are fused using an 8-head multi-head attention mechanism, with initial attention weights set to equal distribution. The model is trained using a cross-entropy loss function, combined with user structured behavior data, with a training batch size of 32 and 50 iterations until the loss value converges.

[0038] In the hyperparameter optimization stage of the model, an evolutionary algorithm was employed, with a genetic algorithm selected as the core evolutionary algorithm. The hyperparameters to be optimized included a learning rate of 0.001 to 0.01, 3 to 7 network layers, 128 to 512 neurons, and an attention weight coefficient of 0.1 to 0.9. The initial hyperparameter population size was 50. The model recommendation accuracy was used as the fitness function. A roulette wheel selection method was used to select high-quality individuals, and a single-point crossover strategy was employed for gene recombination. The crossover probability was set to 0.7, and the adaptive mutation probability was adjusted with the number of iterations. Initially, the probability was 0.3 to explore the parameter space, and later it was reduced to 0.05 to focus on the optimal solution. After 100 generations, the optimal hyperparameter combination was selected: a learning rate of 0.003, 5 network layers, 384 neurons, and an attention weight coefficient of 0.6.

[0039] In the phase of quantifying emotional intensity, a formula is used. Calculation, setting =0.25、 =0.3、 =0.25、 =0.2, satisfying the weighted summation of 1. When the user's heart rate coefficient of variation is normalized... =0.7, Normalized results of skin conductivity fluctuations =0.6, Normalized value of respiratory rate change =0.5, normalized result of body temperature deviation from the baseline value When =0.4, =0.25×0.7+0.3×0.6+0.25×0.5+0.2×0.4=0.175+0.18+0.125+0.08=0.56, the emotional intensity quantification value is 5.6, which is judged as a moderate intensity positive emotion.

[0040] In the dynamic allocation stage of acoustic feature weights, the formula is used. Calculation, setting =0.3, the basic importance weight of tempo, pitch, loudness, timbre spectrum, and melodic contour. The values ​​were 0.2, 0.2, 0.2, 0.2, and 0.2, respectively. The correlation between pace and current emotion under moderate intensity of positive emotion. =0.8, correlation of other acoustic features Both are 0.6, and the final weight of rhythm and speed is calculated. =(1+0.3×0.8)×0.2 / [(1+0.3×0.8)×0.2+4×(1+0.3×0.6)×0.2]≈0.248 / 1.08≈0.23, which is higher than the basic weight, thus achieving the weight adjustment for emotion adaptation.

[0041] In the real-time recommendation list generation stage, the model receives the user's real-time physiological signals and the acoustic characteristics of the currently playing music, outputting an emotional state assessment result of moderate-intensity positive emotion, and calculating the suitability score between the music to be recommended and this emotion. Combining the user's historical preferences for pop and folk styles, a recommendation list of 15 songs is generated by sorting them in descending order of suitability score, and the reasons for suitability are simultaneously displayed on the terminal: the upbeat rhythm matches the current positive mood.

[0042] During the dynamic model update phase, user interaction behaviors related to recommendation results are continuously collected, including playback duration, favorites, sharing, skipping, and replaying. New physiological signals, acoustic features, and interaction data are integrated every 24 hours to incrementally train the model, optimizing feature extraction weights and network parameters. Real-time emotion dynamic tracking uses a default 2-minute sliding time window, calculating an emotion intensity quantification value every 5 seconds. When the change in two consecutive calculations exceeds 2, the recommendation strategy is triggered for real-time updates.

[0043] In the multimodal feedback fusion stage, semantic evaluation text, facial expression images, and operational behaviors from user terminals are collected. The BERT model is used to analyze the sentiment tendency of the text, extract facial key points to identify emotional expressions, and fuse textual, visual, and physiological emotional features with equal weights to correct the model evaluation results. In the recommendation diversity balancing stage, Euclidean distance is used to calculate the acoustic feature similarity of music in the recommendation list. When the average value is higher than 0.7, a diversity adjustment factor of 0.2 is introduced to adjust the ranking results and supplement tracks with significant differences.

[0044] Table 1 Comparison of Music Recommendation Performance in Daily Scenarios on Mobile Terminals Table 1 clearly demonstrates the advantages of this invention in everyday scenarios. Traditional recommendation methods rely on historical behavioral data, resulting in low accuracy in emotion matching, insufficient recommendation diversity, high repetition rates, and limited user satisfaction. This invention, through multi-source data fusion, a dual-branch model, and emotion perception technology, improves emotion matching accuracy by 27 percentage points, significantly increases user satisfaction, greatly enhances recommendation diversity, reduces system response latency by 68.75%, and lowers the repetition rate to 7%. It solves the problem of misalignment between traditional methods and users' real-time moods while ensuring the richness of recommended content and response speed, perfectly adapting to the diverse music consumption needs of everyday scenarios.

[0045] Example 2 In-vehicle intelligent scene intelligent music recommendation application This embodiment is applied to in-vehicle intelligent terminals, adapting to in-vehicle scenarios such as commuting, long-distance travel, and holiday trips. It combines in-vehicle environmental perception with wearable devices to enhance real-time performance and safety, fully implements all technical solutions, and achieves "emotion + behavior driven" music recommendations that are adapted to driving conditions.

[0046] During the multi-source data collection phase, the wearable device wirelessly links with the in-vehicle system, with a sampling frequency set to 15Hz. It collects four physiological signals—heart rate, skin conductivity, respiratory rate, and body temperature—with interference resistance, filtering out signal noise caused by vehicle vibrations, and generating a stable data set every 20 seconds. Using in-vehicle audio analysis tools, the acoustic characteristics of 3000 music tracks adapted to in-vehicle scenarios are extracted from the music library. Tracks with tempos between 80 and 120 BPM and loudness between 50 and 85 dB are prioritized to avoid extreme rhythms or volumes that could affect driving safety. Historical music data from the user's in-vehicle scenarios is simultaneously acquired, including playback records from different driving periods, road condition-related skip actions, and voice command operation records.

[0047] In the preprocessing stage of multi-source features, a moving average method with a window size of 8 is used to denoise physiological signals, enhancing their anti-interference ability. Gaussian filtering is applied to smooth the data, and min-max normalization is performed to the 0-1 range to strictly remove outliers. Acoustic features are Z-score standardized, and structured behavioral data is encoded using one-hot encoding. Feature alignment technology is used to map the data to a 512-dimensional space to construct a fused feature matrix. Audio adaptation preprocessing is added specifically for the in-vehicle environment, adjusting the frequency response curve of the music to match the characteristics of the in-vehicle audio system and optimizing mid-low frequency performance to improve the listening experience in driving scenarios.

[0048] In the dual-branch deep learning model construction phase, the emotion perception branch employs a 4-layer one-dimensional convolutional neural network with kernel sizes of 3, 5, 5, and 7, and the ReLU activation function to enhance the extraction capability of physiological signal features. The acoustic feature branch uses a 2-layer long short-term memory network with 320 hidden units to strengthen the capture of temporal features. The dual-branch features are fused through a 6-head multi-head attention mechanism, and the model is trained using a cross-entropy loss function combined with historical data from in-vehicle scenarios. The training batch size is 48, and the model is iterated for 60 rounds until the loss value converges, ensuring the model's stability in the in-vehicle environment.

[0049] In the hyperparameter optimization stage of the model, an evolutionary algorithm was used. The hyperparameters to be optimized included a learning rate of 0.002 to 0.012, 4 to 8 network layers, 256 to 512 neurons, and an attention weight coefficient of 0.2 to 0.8. The population size was set to 50. The fitness function was a comprehensive indicator of recommendation accuracy and response speed. It was selected through roulette wheel selection, a single-point crossover strategy with a crossover probability of 0.75, and an adaptive mutation probability of 0.35 in the early stage and 0.06 in the later stage. After 100 iterations, the optimal hyperparameters were selected: a learning rate of 0.005, 6 network layers, 420 neurons, and an attention weight coefficient of 0.7.

[0050] In the phase of quantifying emotional intensity, a formula is used. Calculation, setting =0.3、 =0.3、 =0.2、 =0.2. Normalized value of user heart rate coefficient of variation during long-distance driving. =0.5, Normalized results of skin conductivity fluctuations =0.4, Normalized value of respiratory rate change =0.6, Normalized result of body temperature deviation from the baseline value When =0.3, =0.3×0.5+0.3×0.4+0.2×0.6+0.2×0.3=0.15+0.12+0.12+0.06=0.45, the emotional intensity quantification value is 4.5, which is judged as a moderate intensity calm emotion.

[0051] In the dynamic allocation stage of acoustic feature weights, the formula is... At K=0.4, under moderate intensity of calm emotion, the correlation between timbre spectrum and emotion. =0.9, correlation of other acoustic features =0.5, basic importance weight All values ​​are 0.2, resulting in the final weights of the timbre spectrum. =(1+0.4×0.9)×0.2 / [(1+0.4×0.9)×0.2+4×(1+0.4×0.5)×0.2]≈0.272 / 1.08≈0.25, prioritize matching music with a soothing tone.

[0052] In the real-time recommendation list generation phase, the model outputs an emotion assessment result of moderate-intensity calm emotion. Considering the safety requirements of driving scenarios, music with a steady rhythm and moderate volume is selected, and a recommendation list of 15 songs is generated based on suitability scores to avoid highly stimulating tracks that might affect driving. In the dynamic model update phase, incremental training is performed every 12 hours to adapt to driving habits and emotional change patterns.

[0053] In the real-time emotion tracking phase, the sliding time window allows users to adjust it from 1 to 5 minutes, with a default setting of 3 minutes while driving. An emotion intensity quantification is calculated every 5 seconds, and a recommendation update is triggered when the emotion fluctuation exceeds 2, ensuring recommendations are adapted to the driving state. In the multimodal feedback fusion phase, in-vehicle voice evaluation, facial expressions, and user behavior are integrated to refine the recommendation strategy. In the recommendation diversity balancing phase, an adjustment factor of 0.25 ensures diversity while preventing style jumps from affecting driving focus.

[0054] Table 2 Comparison of In-Vehicle Intelligent Scene Music Recommendation Performance Table 2 data highlights the advantages of this invention in in-vehicle scenarios. Traditional recommendation methods suffer from low emotion matching accuracy, insufficient driving safety, lack of recommendation diversity, and high response latency, leading to frequent user operations that impact driving safety. This invention, through in-vehicle scenario-specific optimization, anti-interference data collection, and the combination of emotion perception and safety adaptation, improves emotion matching accuracy by 29 percentage points, significantly enhances driving safety scores, substantially increases recommendation diversity, reduces system response latency by 76.8%, and reduces user operation frequency by 75%. It achieves accurate matching with users' real-time emotions while ensuring driving safety through low operation frequency and safety adaptation design, perfectly meeting the core needs of in-vehicle scenarios.

[0055] Figure 2 This diagram visually demonstrates the core advantages of this invention in emotion perception and matching. Traditional recommendation methods, which do not integrate physiological signals and acoustic features and rely solely on historical behavioral data, achieve an accuracy rate of less than 65% for matching various emotions, particularly with complex emotions such as anger and anxiety. This invention collects physiological signals such as heart rate and skin conductivity through wearable devices, extracts acoustic features such as music rhythm and pitch, and constructs a dual-branch model to establish a precise mapping between emotions and music. The accuracy rate for matching various emotions is improved to over 80%, with a 90% accuracy rate for matching pleasant emotions. This performance confirms the effectiveness of the "emotion + behavior driven" recommendation framework, solving the pain point of misalignment between traditional methods and users' real-time mood, and making recommendation results more aligned with users' needs in different emotional states.

[0056] Figure 3 This figure clearly reflects the supporting role of multi-source feature fusion in recommendation accuracy, highlighting the core technical logic of this invention. Physiological signal features and musical acoustic features contribute a combined 67%, becoming the core driving force for improved recommendation accuracy, confirming the necessity of introducing physiological features and deeply mining unstructured emotional features in music. Historical behavioral features contribute 20%, reflecting the continuity of traditional advantageous features, while the 13% contribution of multimodal feedback features demonstrates the effective supplementation of the closed-loop optimization mechanism. This feature contribution structure breaks away from the traditional recommendation's sole reliance on historical behavioral data. By capturing real-time emotions through physiological signals, matching musical emotions with acoustic features, and dynamically correcting through multimodal feedback, a multi-dimensional and three-dimensional feature support system is constructed, providing a solid foundation for improving recommendation accuracy.

[0057] Figure 4This figure highlights the technical value of this invention in introducing an evolutionary algorithm to optimize hyperparameters. Traditional grid search methods require traversing a large number of parameter combinations, resulting in a recommendation accuracy of only 72%, slow response speed, and numerous convergence iterations, leading to low optimization efficiency. This invention uses a genetic algorithm as the core evolutionary algorithm, rapidly exploring the hyperparameter space through selection, crossover, and mutation operations to select the optimal combination, thereby improving recommendation accuracy by 17 percentage points, reducing system response speed by 72.9%, and decreasing the number of model convergence iterations by 60%. This optimization not only solves the problems of high computational complexity and susceptibility to local optima in traditional hyperparameter tuning but also significantly improves system operating efficiency and real-time performance, providing users with faster and more accurate recommendation services, especially suitable for application scenarios with high response speed requirements, such as mobile terminals and in-vehicle scenarios.

[0058] Figure 5 This diagram comprehensively demonstrates the multi-scenario adaptability and user experience advantages of this invention. Traditional recommendation methods score below 7.5 points in all scenarios, lacking adaptation to user emotional changes and usage needs in different scenarios. For example, in the in-vehicle driving scenario, it does not consider driving safety and emotional stability requirements, and in the sports and fitness scenario, it cannot accurately match the emotional and rhythmic needs during exercise. This invention, through scenario-based data preprocessing, dynamic emotion tracking, and safety adaptation optimization technologies, achieves scores exceeding 8.8 points in all scenarios, reaching 9.3 points in the home leisure scenario. Whether it's the convenience of daily commuting, the safety of in-vehicle driving, or the rhythm adaptability of sports and fitness, it can accurately meet user needs, demonstrating the scenario generalization ability of the "emotion + behavior driven" recommendation framework, solving the problem of insufficient scenario adaptability of traditional recommendation methods, and providing users with a consistent high-quality recommendation experience across all scenarios.

[0059] Figure 6 This diagram visually illustrates the innovation of this invention in recommendation logic, achieving a shift from "behavior-driven" to "emotion + behavior-driven." Traditional recommendation methods (a) rely on 60% of their content on the continuation of historical behavior, with emotion-adaptive recommendations accounting for only 15%, leading to rigid recommended content that is misaligned with real-time moods and causing user fatigue. This invention's method (b) increases the proportion of emotion-adaptive recommendations to 45%, making it the core component of recommended content, while retaining 22% for style preference matching and 8% for diversity supplementation, reducing the proportion of historical behavior continuation to 25%. This composition ensures that recommended content highly matches the user's current mood while also considering long-term style preferences and content richness, breaking the path dependence of traditional recommendations on historical data. Through dynamic emotion tracking and multi-source feature fusion, recommended content becomes more flexible and emotionally resonant, effectively enhancing user engagement and satisfaction, achieving an organic unity of personalization and emotional engagement.

[0060] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent music recommendation based on emotion perception and acoustic features, characterized in that, Comprising the following steps: Collecting multi-source data, collecting physiological signals in real time through user wearable sensing devices, the physiological signals including heart rate, skin conductivity, respiratory rate, body temperature; extracting the acoustic characteristics of the recommended music through audio analysis tools, the acoustic characteristics including rhythm speed, tone height, loudness intensity, timbre spectrum, melody contour; Synchronously acquiring user structured behavior data; Preprocessing multi-source features, denoising, smoothing, and normalizing physiological signals to eliminate abnormal fluctuation data; standardizing acoustic features; Encoding structured behavior data; Mapping the three types of features to the same dimensional space through feature alignment technology to construct a fusion feature matrix; Building a double-branch deep learning model, respectively building an emotion perception branch and an acoustic feature branch; The emotion perception branch uses a convolutional neural network to extract emotion-related features from physiological signals; The acoustic feature branch uses a recurrent neural network to capture the timing association of music acoustic features; Fusing the output features of the double branches through the attention mechanism and training the model parameters in combination with user structured behavior data; Using evolutionary algorithms to optimize model hyperparameters, including learning rate, network layers, neuron number, and attention weight coefficient; initializing the hyperparameter population; taking the model recommendation accuracy as the fitness function; updating the population through selection, crossover, and mutation operations to select the optimal hyperparameter combination and configure it to the model; Generating a real-time recommendation list, the model receives user real-time physiological signals and acoustic features of the current music playing, outputs user real-time emotion evaluation results and music adaptation scores; Based on the evaluation results and scores, a recommendation list containing 10-20 songs is generated; Dynamically updating the model, collecting user real-time feedback behavior, including playing time, collecting, sharing, and skipping; combining new physiological signal and acoustic feature data, periodically incrementally training the model to optimize feature extraction weights and model parameters. 2.The intelligent music recommendation method based on emotion perception and acoustic features according to claim 1, characterized in that, Further comprising an emotion intensity quantification step, combining the time domain and frequency domain features of multiple physiological signals to calculate the real-time emotion intensity of the user, the calculation expression is ; wherein is the emotion intensity quantification value, the value range is 0 to 10; is the heart rate feature weight, is the heart rate variability coefficient normalized value; is the skin conductivity weight, is the skin conductivity fluctuation value normalized result; is the respiration rate weight, is the respiration rate change rate normalized value; is the body temperature weight, is the normalized result of the body temperature deviating from the reference value; . 3.The intelligent music recommendation method based on emotion perception and acoustic features according to claim 1, characterized in that, Further comprising an acoustic feature weight dynamic distribution step, adjusting the contribution weight of each acoustic feature based on the user emotional state and music style preference, the calculation expression is ; wherein is the maximum weight of the acoustic feature of the th class; is the emotional correlation coefficient, the value range is 0.1 to 0.5; is the correlation degree of the acoustic feature of the th class and the current emotion; is the basic importance weight of the acoustic feature of the th class; is the summing operation for all acoustic features. 4.The intelligent music recommendation method based on emotion perception and acoustic features according to claim 1, characterized in that, It also includes a model feature extraction optimization step, introducing an attention gate mechanism in the emotion perception branch, giving different attention weights to physiological signals in different time periods, and focusing on extracting signal features in periods of intense emotional fluctuations; adding a spectrogram attention module in the acoustic feature branch to focus on acoustic frequency band features highly correlated with user emotions; Through the double attention mechanism, the model's ability to capture key features is enhanced, and irrelevant feature interference is reduced. 5.The intelligent music recommendation method based on emotion perception and acoustic features according to claim 1, characterized in that, It also includes a real-time emotion dynamic tracking step, setting a sliding time window to continuously collect physiological signals, with a window length of 30 seconds to 5 minutes and dynamically adjustable; calculating the emotion intensity quantization value every 5 seconds to generate an emotion change curve; when the emotion change amplitude exceeds the preset threshold, triggering the model to update the recommendation strategy in real time, adjusting the acoustic feature adaptation direction, and realizing real-time adjustment of the recommendation result following the user's emotional changes. 6.The intelligent music recommendation method based on emotion perception and acoustic features according to claim 1, characterized in that, It also includes user personalization preference reinforcement steps, builds a user preference vector library, maps user historical behavior data, long-term emotional characteristics, and acoustic feature preferences into high-dimensional vectors; calculates the cosine similarity of new recommended music and preference vectors, combines emotional adaptation scores to generate a comprehensive recommendation index; music with a comprehensive recommendation index higher than the set threshold is preferentially entered into the recommendation list; at the same time, record the user's emotional feedback on different styles of music, and strengthen the extraction and application of personalized preference features. 7.The intelligent music recommendation method based on emotion perception and acoustic features according to claim 1, characterized in that, It also includes a multi-modal feedback fusion step, which collects multi-modal feedback information from the user terminal, including semantic evaluation text, expression images, and operation behaviors; performs sentiment analysis on the semantic evaluation text to extract text emotional features; performs facial feature recognition on the expression images to obtain visual emotional features; and fuses the text emotional features, visual emotional features, and physiological signal emotional features to correct the emotional evaluation results and recommendation strategies of the model, forming a multi-dimensional feedback loop. 8.The intelligent music recommendation method based on emotion perception and acoustic features according to claim 1, characterized in that, It also includes an evolutionary algorithm iterative optimization step, which uses roulette selection to select the operation of the super parameter population; single-point crossover strategy is used for crossover operation; adaptive mutation probability is used for mutation operation, and the mutation probability is dynamically adjusted with the number of iterations. A higher mutation probability is set in the early stage of iteration to explore the super parameter space, and the mutation probability is reduced in the later stage of iteration to focus on the optimal solution; the fitness value of each individual in the population is calculated after each iteration, and the top 30% of individuals with the highest fitness value are retained to enter the next generation population; the iteration stops when the preset number of iterations or the fitness value converges, and the optimal super parameter combination is output. 9.The intelligent music recommendation method based on emotion perception and acoustic features according to claim 1, characterized in that, It also includes a music style subdivision adaptation step, which subdivides music into different subcategories according to style; builds a dedicated acoustic feature template for each style subcategory; calculates the adaptation probability of the current emotion and each style subcategory based on user emotional intensity and style preference; preferentially recommends music in the style subcategory with the highest adaptation probability; and sorts the music in the same style subcategory according to acoustic feature matching degree. 10.The intelligent music recommendation method based on emotion perception and acoustic features according to claim 1, characterized in that, It also includes a recommendation diversity balancing step, which calculates the acoustic feature similarity of different music in the recommendation list; When the average similarity is higher than the set threshold, introduce a diversity adjustment factor to adjust the sorting result, and reduce the sorting priority of highly similar music; select music with large acoustic feature differences from music with higher adaptation degrees to supplement the recommendation list.

Images