An emotion regulation music recommendation method based on fusion of electroencephalogram emotional state and music acoustic characteristics

Abstract

This invention belongs to the field of sentiment analysis technology, specifically relating to a method for recommending music for mood regulation based on the fusion of EEG emotional state and acoustic features of music. The method includes: obtaining discrete emotion category probabilities output by an EEG emotion recognition model and mapping them to continuous two-dimensional emotional coordinates; constructing a user emotional state vector based on the continuous two-dimensional emotional coordinates; obtaining acoustic feature vectors of each candidate music in a candidate music set; inputting historical emotional momentum sequences, temporal context encoding, user emotional state vectors, and acoustic feature vectors into a reward prediction network for processing to obtain the expected mood regulation reward of the candidate music; determining the target recommended music based on the expected mood regulation reward of the candidate music; calculating the actual mood regulation reward after playing the target recommended music to regulate mood and updating the reward prediction network online accordingly. This invention can more precisely describe the user's emotional state, more accurately select music with mood regulation potential, and continuously optimize the recommendation effect through ongoing interaction.

Description

Technical Field

[0001] This invention belongs to the field of emotion analysis technology, specifically relating to a method for recommending music based on emotion regulation by fusing brainwave emotional states with acoustic features of music. Background Technology

[0002] Music, with its strong emotional expression and non-invasive intervention characteristics, has been widely used in scenarios such as mood soothing, stress relief, relaxation training, and digital wellness. However, most existing music recommendation methods are geared towards modeling general interests and preferences, typically recommending music based on users' historical playback, favorites, clicks, or similar user behaviors. This makes it difficult to directly adapt to the characteristics of emotion regulation tasks, such as rapid changes in current emotional state, clear regulation goals, and significant time-varying feedback. In emotion regulation scenarios, matching music based on discrete emotion categories (such as happiness, sadness, etc.) has significant limitations: Firstly, discrete emotion classification results can only provide limited category labels, making it difficult to characterize fine-grained differences in the intensity, direction, and trend of emotional states. Furthermore, users' music preferences and emotional response patterns vary across different time periods. Secondly, music itself possesses multidimensional acoustic properties; different rhythms, energy levels, spectral structures, harmonic structures, and timbre characteristics can produce varying regulatory effects on different emotional states. In existing technologies, few solutions can uniformly model EEG emotion recognition results, continuous emotional states, historical emotional change trends, temporal context, and musical acoustic features. Furthermore, there is a lack of a mechanism that can continuously receive feedback and adjust recommendation strategies online during the recommendation process.

[0003] In conclusion, there is an urgent need for an emotion-modulating music recommendation method that can integrate and model real brainwave emotions with the acoustic features of music, in order to improve the targeting, dynamic adaptability and personalization of music recommendations. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention proposes a method for recommending emotion-modulating music based on the fusion of EEG emotional states and musical acoustic features. This method includes:

[0005] S1: Obtain the discrete emotion category probabilities output by the EEG emotion recognition model and map them into continuous two-dimensional emotion coordinates;

[0006] S2: Construct emotional momentum based on continuous two-dimensional emotional coordinates; encode the current time to obtain time context encoding; concatenate the continuous two-dimensional emotional coordinates, emotional momentum, and time context encoding to obtain the user's emotional state vector;

[0007] S3: Obtain the acoustic feature vector of each candidate music in the candidate music set;

[0008] S4: Obtain the emotional momentum at historical moments and construct a historical emotional momentum sequence; input the historical emotional momentum sequence, temporal context encoding, user emotional state vector, and acoustic feature vector into the reward prediction network for processing to obtain the expected modulated reward of the candidate music;

[0009] S5: Adjust rewards based on the expected performance of candidate music to determine target recommended music;

[0010] S6: Calculate the actual adjustment reward after playing the target recommended music to regulate mood; calculate the predicted total loss based on the actual adjustment reward and the expected adjustment reward of the target recommended music, and update the reward prediction network online based on the predicted total loss.

[0011] Preferably, the discrete emotion category probabilities are mapped to continuous two-dimensional emotion coordinates as follows:

[0012]

[0013] in, This represents the continuous two-dimensional sentiment coordinates at the current time t. Let represent the probability of the i-th emotion. This represents the center point of the i-th quadrant.

[0014] Preferably, the acoustic feature vector includes rhythm waveform (BPM), mean and variance of energy (RMS), mean and variance of spectral centroid, chroma feature, and Mel frequency cepstral coefficient (MFCC) feature.

[0015] Preferably, the process of obtaining the expected adjusted reward for candidate music includes:

[0016] A temporal context encoding module is used to process the temporal context encoding to obtain temporal context embedding features; the temporal context encoding module consists of a first fully connected layer, an activation function, a second fully connected layer, and a normalization layer connected sequentially;

[0017] A memory enhancement module is used to process the historical emotional momentum sequence to obtain a memory-enhanced representation; the memory enhancement module is implemented using a gated recurrent unit in a recurrent neural network.

[0018] A cross-modal attention module is used to process the user's emotional state vector and acoustic feature vector to obtain cross-modal fusion features;

[0019] Multi-source features are obtained by concatenating temporal context embedding features, memory-enhanced representations, and cross-modal fusion features.

[0020] Multi-source features are input into the fusion prediction module for processing to obtain the expected adjusted reward for candidate music.

[0021] Furthermore, the cross-modal attention module processes the user's emotional state vector and acoustic feature vector, including:

[0022] A query matrix is ​​obtained by linearly mapping the user's emotional state vector; different linear mappings are performed on the acoustic feature vectors to obtain the key matrix and value matrix.

[0023] Calculate the standard attention output based on the query matrix, key matrix, and value matrix;

[0024] The standard attention output and the user's emotional state vector are concatenated, and a gating value is constructed based on the concatenation result;

[0025] The cross-modal fusion feature is obtained by multiplying the gating value element-wise with the standard attention output.

[0026] Preferably, the formula for determining the target recommended music based on the expected adjustment of rewards for candidate music is:

[0027]

[0028]

[0029]

[0030] in, This represents the target music recommendation at the current time t. Indicates the candidate music at the current time t. The overall score, Indicates the candidate music at the current time t. Expectation-adjusted rewards Indicates the exploration coefficient. Indicates the candidate music at the current time t. Exploration items, This indicates the candidate music up to the current time t. The number of times it was selected.

[0031] Preferably, the process of calculating the actual adjustment reward includes:

[0032] Calculate the magnitude of emotional changes after playing target-recommended music;

[0033] Calculate the cosine similarity between the user's actual regulatory state change and the target regulatory state change based on the amount of emotional change after playing the target recommended music;

[0034] The actual moderated reward is calculated based on the magnitude of emotional change and cosine similarity.

[0035] Furthermore, the formula for calculating the cosine similarity between the user's actual adjustment state change and the target adjustment state change is:

[0036]

[0037] in, This represents the cosine similarity between the user's actual adjustment state change and the target adjustment state change at the current time t. This indicates the magnitude of emotional change after playing the recommended music. This indicates the preset target emotional point. This represents the continuous two-dimensional sentiment coordinates at the current time t. This represents the L2 norm.

[0038] Furthermore, the formula for calculating the actual moderated reward based on the magnitude of emotional change and cosine similarity is as follows:

[0039]

[0040] in, This represents the actual adjusted reward at the current time t. This represents the cosine similarity between the user's actual adjustment state change and the target adjustment state change at the current time t. and This indicates the emotional impact of playing recommended music compared to playing it before. Indicates the weighting coefficient for improving effectiveness value. This represents the consistency weighting coefficient for the target direction. This represents the weighting coefficient for the magnitude of emotional changes. This represents the L2 norm.

[0041] Preferably, the formula for calculating the total predicted loss is:

[0042]

[0043] in, express, Represents the total number of samples. This indicates that the expected reward for the candidate music in the i-th sample of the reward prediction network output is adjusted. This represents the multi-source features of the i-th sample within the experience buffer. This represents the actual adjusted reward at the current time t.

[0044] The beneficial effects of this invention are as follows:

[0045] This invention directly utilizes the discrete emotion category probabilities output by an EEG emotion recognition model employing deep learning methods. Instead of relying on the rigid classification of traditional discrete labels, it achieves a fine-grained description of the user's emotional position through continuous two-dimensional emotional coordinate soft mapping, thereby enhancing the ability of recommendation decisions to express the current emotional state. By introducing emotional momentum changes to explicitly represent the direction and magnitude of user emotional changes, the recommendation system not only focuses on the current emotional position but also perceives the trend of emotional evolution. The designed temporal context encoding module enables the recommendation algorithm to perceive changes in user preferences over different time periods and adapt to differences in adjustment needs in a timely manner. The memory enhancement module proposed in this invention models historical emotional momentum sequences, enabling rapid determination of the user's current potential adjustment response pattern and individual preference characteristics, significantly reducing cold start time compared to traditional recommendation algorithms. The cross-modal attention module establishes a correlation between the user's emotional state and the acoustic features of the music, allowing the system to select more suitable acoustic attributes for different emotional states, improving the targeting of music adjustment. This invention improves the personalization, dynamic adaptability, and adjustment effectiveness of music recommendations in emotion adjustment scenarios. Attached Figure Description

[0046] Figure 1 This is a diagram illustrating the overall framework of the emotion regulation music recommendation method in this invention.

[0047] Figure 2 This is a schematic diagram of the continuous two-dimensional valence-arousal coordinate system in this invention;

[0048] Figure 3 This is a schematic diagram of the TCE module in this invention;

[0049] Figure 4 This is a schematic diagram of the MA module in this invention;

[0050] Figure 5 This is a schematic diagram of the CMAM module in this invention. Detailed Implementation

[0051] 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.

[0052] This invention proposes a music recommendation method for emotion regulation based on the fusion of EEG emotional state and acoustic features of music. First, the discrete emotion category probabilities output by the EEG emotion recognition model are transferred to a continuous valence-arousal two-dimensional space to obtain a more granular current emotion position. Then, a user's emotional state vector is constructed by combining emotion change trends and temporal context. Next, this user's emotional state vector is fused with the acoustic features of candidate music across modalities to predict the moderating effect of the candidate music. Finally, a NeuralUCB online decision-making strategy with an exploration mechanism is used to select target music and continuously iterate and update the recommendation strategy. Figure 1 As shown, the method specifically includes the following:

[0053] S1: Obtain the discrete emotion category probabilities output by the EEG emotion recognition model and map them into continuous two-dimensional emotion coordinates.

[0054] The EEG emotion recognition model outputs the Softmax probabilities corresponding to four emotion categories (happiness, fear, sadness, and calmness), which respectively correspond to four quadrants: high valence and high arousal, low valence and high arousal, low valence and low arousal, and high valence and low arousal. Figure 2 As shown, the continuous two-dimensional valence-arousal coordinate axis ranges from [-1, 1]. The center points of the four quadrants are set as (0.7, 0.7), (-0.7, 0.7), (-0.7, -0.7), and (0.7, -0.7), respectively.

[0055] The four emotion probabilities output by the EEG emotion recognition model are represented as p1, p2, p3, and p4, respectively, and the corresponding center points of the four quadrants are represented as C1, C2, C3, and C4, respectively.

[0056] By using a probability-weighted centroid approach to achieve continuous coordinate mapping, the current continuous two-dimensional emotional coordinate point P... t It can be expressed by the following formula:

[0057]

[0058] in, P represents the continuous two-dimensional emotion coordinates at the current time t. t =(v t , a t ), v t a represents the effective value at time t. t This represents the wake-up value at time t. Let represent the probability of the i-th emotion. This represents the center point of the i-th quadrant. Using the above method, the originally discrete emotion recognition results are transformed into continuous two-dimensional emotion locations, which is beneficial for subsequent refined adjustment modeling.

[0059] S2: Construct emotional momentum based on continuous two-dimensional emotional coordinates; encode the current time to obtain time context encoding; concatenate the continuous two-dimensional emotional coordinates, emotional momentum and time context encoding to obtain the user's emotional state vector.

[0060] Emotional momentum is constructed based on continuous two-dimensional emotional coordinates. Specifically, the trend of emotional change is represented by the emotional momentum vector. accomplish, We obtain it from the following formula:

[0061]

[0062] in, Let represent the emotional momentum at the current time t, with a dimension of 2; This represents the continuous two-dimensional emotional coordinates from the previous moment. Emotional momentum vector. It is used to reflect the direction and magnitude of changes in user emotions over a continuous period of time.

[0063] Encode the current time to obtain a time context code. Specifically, divide the day into four time periods: morning, afternoon, evening, and night. Generate a one-hot encoded time context code based on the time period to which the current time belongs. .

[0064] By concatenating continuous two-dimensional emotion coordinates, emotion momentum, and temporal context encoding, a user emotion state vector is obtained; the user emotion state vector is represented as follows:

[0065]

[0066] in, This represents the user's emotional state vector at the current time t. This represents the continuous two-dimensional sentiment coordinates at the current time t. This represents the emotional momentum at the current time t. This represents the time context code for the current time t.

[0067] S3: Obtain the acoustic feature vector of each candidate music in the candidate music set.

[0068] Obtain the acoustic feature vector of each candidate music in the candidate music set. This is used for subsequent cross-modal fusion. The musical acoustic features include rhythm waveform (BPM), energy RMS mean and variance, spectral centroid mean and variance, chroma features, and Mel-frequency cepstral coefficients (MFCC) features, forming a unified musical acoustic feature vector. Specifically, the musical acoustic feature vector has a dimension of 30, with BPM being 1-dimensional, RMS 2-dimensional, spectral centroid 2-dimensional, chroma features 12-dimensional, and MFCC 13-dimensional.

[0069] S4: Obtain the emotional momentum at historical moments and construct a historical emotional momentum sequence; input the historical emotional momentum sequence, temporal context encoding, user emotional state vector, and acoustic feature vector into the reward prediction network for processing to obtain the expected modulated reward of the candidate music.

[0070] Obtain the emotional momentum at historical moments and construct a historical emotional momentum sequence, which is represented as follows: ,in, This is the length of the history window.

[0071] The reward prediction network designed in this invention includes a temporal context encoding module (TCE), a memory enhancement module (MA), a cross-modal attention module (CMAM), and a fusion prediction module.

[0072] The process of processing historical sentiment momentum sequences, temporal context encoding, user sentiment state vectors, and acoustic feature vectors in the reward prediction network includes:

[0073] The temporal context encoding module is used to process the temporal context encoding to obtain temporal context embedding features, specifically:

[0074] like Figure 3 As shown, the TCE module consists of a first fully connected layer, an activation function, a second fully connected layer, and a normalization layer connected sequentially. It is used to map one-hot temporal encodings to a continuous embedding space. The input is the temporal encoding, and the output is the temporal context embedding feature. From the following formula, we can obtain:

[0075]

[0076] in, T is the ReLU activation function. t = [ , , , ] represents the context code of the time period to which the current time belongs, W t1 W t2 b is a learnable weight matrix t1 b t2 For bias terms, For intermediate output features, This indicates a normalization operation.

[0077] The historical emotional momentum sequence is processed using a memory enhancement module to obtain a memory-enhanced representation, specifically:

[0078] like Figure 4 As shown, the MA module is implemented using a gated recurrent unit (GRU) in a recurrent neural network, and the historical sentiment momentum sequence is... Where k is the historical window length. The normalized hidden state at the last moment is taken. As a representation of memory enhancement Satisfy the following formula:

[0079]

[0080] This MA module is used to model historical emotional momentum sequences, extract users' implicit regulatory preferences during a continuous interaction process, and understand individual regulatory habits, short-term response style patterns, and potential preferences.

[0081] A cross-modal attention module is used to process the user's emotional state vector and acoustic feature vector to obtain cross-modal fusion features, specifically:

[0082] like Figure 5 As shown, the user's emotional state vector S t The acoustic feature vector x of the a-th candidate music in the candidate music set a Perform linear mappings separately:

[0083]

[0084] Among them, Q t K a V a W represents the query matrix, key matrix, and value matrix, respectively. Q W K W V All of these are learnable parameter matrices.

[0085] Calculate the standard attention output based on the query matrix, key matrix, and value matrix:

[0086]

[0087] in, For standard attention output, It is the dimension of the key matrix.

[0088] Concatenate the standard attention output and the user's emotional state vector, and construct a gating value based on the concatenation result:

[0089]

[0090] in, Indicates the gate value, This indicates the output of attention. and user emotional state vector S t To splice, W is the Sigmoid activation function. g With b g All of these are gating parameters.

[0091] Multiplying the gate value element-wise with the standard attention output yields the cross-modal fusion feature:

[0092]

[0093] in, Indicates the candidate music at the current time t. Cross-modal fusion features, This indicates element-wise multiplication.

[0094] This CMAM module is used to establish a cross-modal association between the current user's emotional state and the acoustic features of candidate music. Based on the current emotional state, it dynamically focuses on the key acoustic attributes of different music to achieve the goal of selecting which type of music to bring about positive emotional regulation in the current state.

[0095] splicing temporal context embedding features Memory enhancement representation and cross-modal fusion features Multi-source features are obtained:

[0096]

[0097] Multi-source features are input into the fusion prediction module for processing to obtain the expected adjusted reward for candidate music. Specifically, it satisfies the following formula:

[0098]

[0099] in, W represents a reward prediction network. o W f For learnable parameter matrix, b is the ReLU activation function. f b o This is a bias term.

[0100] S5: Adjust rewards based on the expected performance of candidate music to determine target recommended music.

[0101] After obtaining the expected adjusted reward for all candidate music tracks, the NeuralUCB algorithm with an exploration-utilization balance mechanism is used to select the target music. The overall score for candidate music track 'a' is calculated. Defined as:

[0102]

[0103] In the formula The exploration coefficient, greater than 0, is used to indicate the strength of exploration ability. For exploration, the term is preferably:

[0104]

[0105] Among them This indicates the number of times candidate music a has been selected up to the current time t, i.e., in round t.

[0106] The music ultimately recommended is:

[0107]

[0108] The reward prediction network is used to predict the reward value for each piece of music. The exploration in the NeuralUCB algorithm utilizes a balancing mechanism, which allows the exploration term to fully utilize the optimal adjustment music features learned by the network while retaining the ability to try music that has not been fully explored, thus avoiding getting stuck in local optima and causing excessive recommendation of duplicate content.

[0109] S6: Calculate the actual adjustment reward after playing the target recommended music to regulate mood; calculate the predicted total loss based on the actual adjustment reward and the expected adjustment reward of the target recommended music, and update the reward prediction network online based on the predicted total loss.

[0110] After the target music is recommended and played, the user's emotion drifts along a preset emotional point or in the opposite direction, and after being superimposed with noise, a new emotional state is formed. This invention calculates the actual adjustment reward based on user feedback, i.e., the updated emotional state change result, specifically:

[0111] The change in mood after playing recommended music can be expressed as:

[0112]

[0113] in, The user's emotional state before the recommendation. This represents the user's emotional state after the recommendation.

[0114] The cosine similarity between the user's actual adjustment state change and the target adjustment state change satisfies the following formula:

[0115]

[0116] in, This represents the cosine similarity between the user's actual adjustment state change and the target adjustment state change at the current time t. This indicates the magnitude of emotional change after playing the recommended music. This indicates the preset target emotional point, which is set within the high valence-high arousal quadrant; Represents the continuous two-dimensional sentiment coordinates at the current time t; This represents the L2 norm, which reflects the consistency between the actual direction of emotional change and the target direction.

[0117] The actual moderated reward is calculated based on the magnitude of emotional change and cosine similarity.

[0118]

[0119] in, This represents the actual adjusted reward at the current time t. This represents the cosine similarity between the user's actual adjustment state change and the target adjustment state change at the current time t. and This indicates the emotional impact of playing recommended music compared to playing it before. Indicates the weighting coefficient for improving effectiveness value. This represents the consistency weighting coefficient for the target direction. This represents the weighting coefficients for the magnitude of emotion changes. The first term measures the improvement in effectiveness, the second term measures the similarity between the direction of emotion change and the target direction, and the third term measures the magnitude of emotion regulation changes.

[0120] The calculated actual adjustment reward can be transformed into a supervision signal for algorithm updates, allowing the algorithm to further explore which music genres cause more significant changes in emotional state. The method of this invention updates the reward prediction network online based on the actual adjustment reward obtained. Specifically:

[0121] The samples obtained during the recommendation process are written into the NeuralUCB algorithm's experience buffer. These samples include state vectors, selected music, modulated rewards, historical sentiment momentum sequences, and temporal features. Once a preset number of samples are obtained, batches are extracted for training. The total loss function is predicted. The following calculations were made based on the actual and expected adjustment rewards of the target recommended music:

[0122]

[0123] in, Represents the total number of samples. This indicates that the expected reward for the candidate music in the i-th sample of the reward prediction network output is adjusted. This represents the multi-source features of the i-th sample within the experience buffer. This represents the actual adjusted reward at the current time t.

[0124] The parameter update process is as follows:

[0125]

[0126] In the formula For all learnable parameters of the network, For learning rate, Indicates the loss with respect to parameters The direction of fastest gradient descent.

[0127] This operation can continuously refine the reward prediction network by utilizing real feedback after recommendations, enabling the recommendation algorithm to gradually adapt to different users and real-world scenarios during continuous interaction.

[0128] In summary, this invention proposes a music recommendation method for emotion regulation based on the fusion of EEG emotional state and musical acoustic features. This method starts with the discrete probability output of an EEG emotion recognition model, obtains fine-grained emotion locations through continuous two-dimensional emotion mapping, constructs a state vector by combining emotional momentum and temporal context, and achieves joint modeling of temporal features, historical dynamics, and musical acoustic features through three modules: TCE, MA, and CMAM. Finally, it employs an online decision-making strategy based on exploratory utilization balance (NeuralUCB) to output target music and continuously updates the recommendation strategy based on feedback. Compared with existing technologies, this invention can more precisely describe the user's emotional state, more accurately select music with regulation potential, and continuously optimize the recommendation effect through continuous interaction, demonstrating high practical value and promotional significance.

[0129] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for recommending music based on the fusion of brainwave emotional states and musical acoustic features, characterized in that, include: S1: Obtain the discrete emotion category probabilities output by the EEG emotion recognition model and map them into continuous two-dimensional emotion coordinates; S2: Construct emotional momentum based on continuous two-dimensional emotional coordinates; Encode the current time to obtain the time context code; By concatenating continuous two-dimensional emotional coordinates, emotional momentum, and temporal context encoding, a user's emotional state vector is obtained. S3: Obtain the acoustic feature vector of each candidate music in the candidate music set; S4: Obtain the emotional momentum at historical moments and construct a historical emotional momentum sequence; input the historical emotional momentum sequence, temporal context encoding, user emotional state vector, and acoustic feature vector into the reward prediction network for processing to obtain the expected modulated reward of the candidate music; S5: Adjust rewards based on the expected performance of candidate music to determine target recommended music; S6: Calculate the actual adjustment reward after playing the target recommended music to regulate mood; calculate the predicted total loss based on the actual adjustment reward and the expected adjustment reward of the target recommended music, and update the reward prediction network online based on the predicted total loss.

2. The method for recommending emotion-regulating music based on the fusion of EEG emotional state and musical acoustic features according to claim 1, characterized in that, The discrete emotion category probability is mapped to a continuous two-dimensional emotion coordinate as follows: ； in, This represents the continuous two-dimensional sentiment coordinates at the current time t. Let represent the probability of the i-th emotion. This represents the center point of the i-th quadrant.

3. The method for recommending emotion-regulating music based on the fusion of EEG emotional state and musical acoustic features as described in claim 1, characterized in that, The acoustic feature vector includes rhythm waveform (BPM), mean and variance of energy (RMS), mean and variance of spectral centroid, chroma feature, and Mel frequency cepstral coefficient (MFCC) feature.

4. The method for recommending emotion-regulating music based on the fusion of EEG emotional state and musical acoustic features according to claim 1, characterized in that, The process of obtaining the expected adjusted reward for candidate music includes: A temporal context encoding module is used to process the temporal context encoding to obtain temporal context embedding features; the temporal context encoding module consists of a first fully connected layer, an activation function, a second fully connected layer, and a normalization layer connected sequentially; A memory enhancement module is used to process the historical emotional momentum sequence to obtain a memory-enhanced representation; the memory enhancement module is implemented using a gated recurrent unit in a recurrent neural network. A cross-modal attention module is used to process the user's emotional state vector and acoustic feature vector to obtain cross-modal fusion features; Multi-source features are obtained by concatenating temporal context embedding features, memory-enhanced representations, and cross-modal fusion features. Multi-source features are input into the fusion prediction module for processing to obtain the expected adjusted reward for candidate music.

5. The method for recommending emotion-regulating music based on the fusion of EEG emotional state and musical acoustic features according to claim 4, characterized in that, The process by which the cross-modal attention module processes the user's emotional state vector and acoustic feature vector includes: A query matrix is ​​obtained by linearly mapping the user's emotional state vector; different linear mappings are performed on the acoustic feature vectors to obtain the key matrix and value matrix. Calculate the standard attention output based on the query matrix, key matrix, and value matrix; The standard attention output and the user's emotional state vector are concatenated, and a gating value is constructed based on the concatenation result; The cross-modal fusion feature is obtained by multiplying the gating value element-wise with the standard attention output.

6. The method for recommending emotion-regulating music based on the fusion of EEG emotional state and musical acoustic features according to claim 1, characterized in that, The formula for determining the target recommended music based on the expected reward adjustment of candidate music is: ； ； ； in, This represents the target music recommendation at the current time t. Indicates the candidate music at the current time t. The overall score, Indicates the candidate music at the current time t. Expectation-adjusted rewards Indicates the exploration coefficient. Indicates the candidate music at the current time t. Exploration items, This indicates the candidate music up to the current time t. The number of times it was selected.

7. The method for recommending emotion-regulating music based on the fusion of EEG emotional state and musical acoustic features according to claim 1, characterized in that, The process of calculating the actual adjusted reward includes: Calculate the magnitude of emotional changes after playing target-recommended music; Calculate the cosine similarity between the user's actual regulatory state change and the target regulatory state change based on the amount of emotional change after playing the target recommended music; The actual moderated reward is calculated based on the magnitude of emotional change and cosine similarity.

8. The method for recommending emotion-regulating music based on the fusion of EEG emotional state and musical acoustic features according to claim 7, characterized in that, The formula for calculating the cosine similarity between the user's actual change in accommodation state and the target change in accommodation state is: ； in, This represents the cosine similarity between the user's actual adjustment state change and the target adjustment state change at the current time t. This indicates the magnitude of emotional change after playing the recommended music. This indicates the preset target emotional point. This represents the continuous two-dimensional sentiment coordinates at the current time t. This represents the L2 norm.

9. The method for recommending emotion-regulating music based on the fusion of EEG emotional state and musical acoustic features according to claim 7, characterized in that, The formula for calculating the actual moderated reward based on the magnitude of emotional change and cosine similarity is as follows: ； in, This represents the actual adjusted reward at the current time t. This represents the cosine similarity between the user's actual adjustment state change and the target adjustment state change at the current time t. and This indicates the emotional impact of playing recommended music compared to playing it before. Indicates the weighting coefficient for improving effectiveness value. This represents the consistency weighting coefficient for the target direction. This represents the weighting coefficient for the magnitude of emotional changes. This represents the L2 norm.

10. The method for recommending emotion-regulating music based on the fusion of EEG emotional state and musical acoustic features according to claim 1, characterized in that, The formula for calculating the total predicted loss is: ； in, express, Represents the total number of samples. This indicates that the expected reward for the candidate music in the i-th sample of the reward prediction network output is adjusted. This represents the multi-source features of the i-th sample within the experience buffer. This represents the actual adjusted reward at the current time t.

Images