Intelligent music recommendation methods and systems based on user emotion regulation
By estimating the user's current emotional state and generating the target emotional state, and combining music VA and tag prediction attributes, this approach solves the problem of continuous emotion regulation in existing music recommendation systems, achieving higher emotion adaptability and semantic interpretability in music recommendations.
Patent Information
- Application Number
- CN202611189975.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-06
- Publication Date
- 2026-11-17
AI Technical Summary
Existing music recommendation systems struggle to continuously adjust emotions based on users' real-time emotional states and lack a joint recommendation path that integrates user textual emotions, music's continuous emotional attributes, and music's semantic tags.
By acquiring user input text, the current emotional state is estimated and the target emotional state is generated. The music VA prediction attribute and music Tag prediction attribute of candidate music are combined to calculate a joint ranking function for music recommendation.
It enables intelligent music recommendation that continuously adjusts emotions based on the user's real-time emotional state, improving the emotional adaptability and semantic interpretability of the recommendation results.
Smart Images

Figure CN123115457A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to an intelligent music recommendation method and system for user emotion regulation. Background Technology
[0002] With the development of digital music platforms, mobile devices, and online audio services, users can easily access a vast amount of music content. Existing music recommendation systems typically recommend music to users based on their historical playback behavior, favorites, collaborative filtering results, song similarity, audio content characteristics, or manually labeled mood tags.
[0003] However, in the context of mood regulation music recommendation, users typically need more than just songs similar to their historical preferences. They want music that can help them maintain their current emotional state, calm them down, provide comfort, or uplift them. Existing recommendation methods based on user history primarily reflect long-term preferences and struggle to determine the current emotional state from the user's immediate natural language input. Existing recommendation methods based on discrete emotion tags usually label music into several fixed emotion categories, making it difficult to depict the continuous changes between the current and target emotional states. While existing recommendation methods based on audio features or song similarity can measure music content similarity to some extent, they generally lack an interpretable recommendation path that combines user textual emotion, continuous emotional attributes of the music, and semantic tags of the music.
[0004] Furthermore, in practical music recommendation systems, candidate music often lacks stable and complete artificial emotion labels. Relying solely on artificial labels or single emotion categories for recommendations can easily lead to a lack of adjustment direction in a continuous emotion space, and it's difficult to explain the relationship between the recommended music and the user's adjustment goals. Therefore, an intelligent music recommendation method is needed that can estimate the current emotional state based on user text input, generate the target emotional state based on adjustment strategies and intensity, and perform joint ranking by combining music VA prediction attributes and music tag prediction attributes. Summary of the Invention
[0005] The purpose of this invention is to provide an intelligent music recommendation method based on user emotion regulation, overcoming the shortcomings of traditional recommendation methods that struggle to continuously regulate recommendations based on the user's real-time emotional state, and offering the advantage of high emotion adaptability in the recommendation results. The technical solution adopted is as follows: A smart music recommendation method based on user emotion regulation includes the following steps: Step 1: Obtain user input text, adjustment strategy, adjustment intensity, and candidate music set; Step 2: Estimate the current emotional state based on the user's input text. The current emotional state The coordinates are in the two-dimensional continuous emotion space of valence-arousal. Step 3: Based on the current emotional state Regulation strategies and regulation intensity generate target emotional states Among them, strategy anchor point Based on the Tag prototype corresponding to the preset representative label Confirmed, the Tag prototype These are fixed values pre-stored in the database; Step 4: Obtain the music VA prediction attribute for each candidate music in the candidate music set. The music VA prediction attribute The coordinates of the candidate music in the valence-arousal two-dimensional continuous emotion space; Step 5: Obtain the music tag prediction attribute for each candidate music in the candidate music set. The music tag prediction attribute The multidimensional probability vector is formed by the predicted probabilities of the candidate music on each tag in the preset tag system; Step 6: Based on the target emotional state and the pre-stored prototypes of each of the aforementioned tags Generate target tag distribution ; Step 7: Based on the current emotional state The target emotional state The music VA prediction attribute The music tag prediction attribute and the target tag distribution Calculate the joint ranking function for each candidate music. The joint sorting function This includes target emotion proximity items, current emotion association items, and tag semantic matching items; Step 8: According to the joint sorting function Sort the candidate music in the candidate music set and output a Top-K music recommendation list.
[0006] Preferably, in step 3, based on the current emotional state... The regulation strategy and the regulation intensity generate the target emotional state. ,include: Determine the strategy anchor point based on the aforementioned adjustment strategy. Through formula Calculate the target emotional state, wherein, The adjustment intensity is defined as having a value range of [0,1]. The regulation strategy includes at least one of the following: maintaining the current state strategy, calming strategy, comforting strategy, and cheering strategy; When the adjustment strategy is to maintain the current state, the strategy anchor point When the regulation strategy is a calming strategy, a comforting strategy, or an uplifting strategy, the strategy anchor point... Tag prototypes representing the corresponding strategies Sure.
[0007] Preferably, the Tag prototype Pre-build using the following methods: Step A: Obtain the music VA prediction attribute for each candidate music in the candidate music set. Music Tag Prediction Attribute ; Step B: For each type of label in the preset label system According to the candidate music in the tags On the predicted probability Sort the candidate music from highest to lowest, and select the top 5% as the high-confidence sample set. ; Step C: Using the predicted probability of the high-confidence sample As weights, the music VA prediction attributes of the high-confidence samples are... Perform a weighted average to obtain the label. Tag prototype .
[0008] Preferably, the music VA prediction attribute and the music tag prediction attribute The fixed values are pre-calculated and stored in the database through audio feature extraction and model prediction; the online recommendation stage queries and retrieves these values from the database. and stated .
[0009] Preferably, in step 6, based on the target emotional state and the pre-stored prototypes of each of the aforementioned tags Generate target tag distribution ,include: Step 6A: Calculate the target emotional state Tag prototypes for each type of label Euclidean distance in valence-arousal space; Step 6B: Convert the Euclidean distance into an activation value using a negative exponential mapping function; Step 6C: Perform Softmax normalization on the activation values of all tags to obtain the target tag distribution.
[0010] Preferably, the joint sorting function in step 7 include: Target Emotion Proximity Item The music VA prediction attribute used to measure the candidate music. With the target emotional state The degree of closeness; Current sentiment related items The music VA prediction attribute used to measure the candidate music. With the current emotional state The degree of connection; Tag semantic matching items The music tag prediction attribute used to measure the candidate music. Distribution of the target tag The degree of matching; The joint sorting function is: ,in , , For preset weight parameters, and .
[0011] An intelligent music recommendation system based on user emotion regulation includes: The user interface module is used to obtain user input text, adjustment strategy, adjustment intensity, and candidate music set; The text sentiment estimation module is used to estimate the current sentiment state based on the user's input text. The current emotional state The coordinates are in the two-dimensional continuous emotion space of valence-arousal. The target emotion state generation module is used to generate emotions based on the current emotion state. The regulation strategy and the regulation intensity generate the target emotional state. ; A database is used to pre-store the music VA prediction attributes of each candidate music in the candidate music set. Music Tag Prediction Attributes And the tag prototype for each type of tag in the preset tag system. ; The audio feature extraction module is used to extract audio features from candidate music tracks; The music VA prediction module is used to output the music VA prediction attribute of the candidate music based on the audio features. ; The music tag prediction module is used to output the music tag prediction attribute of the candidate music based on the audio features. ; Tag prototype building module, used to predict attributes based on the music VA. and the music tag prediction attribute For each type of tag in the preset tag system, construct a tag prototype. ; The target tag distribution generation module is used to generate tags based on the target emotional state. and the pre-stored prototypes of each of the aforementioned tags Generate target tag distribution ; The joint sorting module is used to sort according to the current emotional state. The target emotional state The music VA prediction attribute The music tag prediction attribute and the target tag distribution Calculate the joint ranking function for each candidate music. and according to the joint sorting function Sort the candidate music tracks and output a Top-K music recommendation list; And a recommendation explanation information output module, used to output recommendation explanation information.
[0012] Compared with the prior art, the advantages of the present invention are: 1. This invention provides an intelligent music recommendation method for user emotion regulation, which can integrate user input text, regulation strategy, regulation intensity and candidate music set into the recommendation process. First, the current emotional state is estimated based on the user input text, then the target emotional state is generated based on the current emotional state, regulation strategy and regulation intensity, and finally the joint ranking recommendation is performed based on the music VA prediction attribute and music Tag prediction attribute of the candidate music.
[0013] Compared with recommendation methods that rely solely on users' historical playback behavior, song similarity, or discrete emotion tags, this invention not only considers users' long-term preferences or song content similarity, but also determines the current emotional state based on the user's real-time text input, and generates corresponding target emotional states based on different adjustment strategies such as maintaining the current state, calming, comforting, and cheering up, thus making it more suitable for emotion regulation scenarios.
[0014] 2. This invention unifies the user's current emotional state, target emotional state, and the music VA prediction attributes of candidate music into a two-dimensional continuous emotional space of valence-arousal. This allows for the depiction of the continuous changing relationship between the current and target emotional states, avoiding the limitation of traditional discrete emotional tag recommendation methods that can only perform fixed-category matching. Through target emotional proximity and current emotional association, this invention can select candidate music that is closer to the target emotional state while maintaining a smooth connection between the candidate music and the user's current emotional tone, resulting in a more explicit direction for emotional regulation in the recommendation results.
[0015] 3. This invention does not require candidate music to have real artificial emotion labels, but uses music VA prediction attributes and music tag prediction attributes as model prediction attributes or pseudo attributes.
[0016] High-confidence samples are selected by model prediction probabilities of candidate music on each music tag, and the music VA prediction attributes of the high-confidence samples are weighted and averaged to construct tag prototypes, so that discrete tag semantics can be mapped to continuous VA sentiment space.
[0017] Therefore, this invention can still achieve emotion-semantic joint recommendation using model prediction results even in real-world music library scenarios where manual labels are incomplete, unstable, or costly.
[0018] 4. This invention generates a target tag distribution based on the distance between the target emotional state and each tag prototype, giving higher weights to tag prototypes closer to the target emotional state. Through this mechanism, this invention can further transform the "target emotional state" into tag semantic preferences that can be used for music ranking, making the recommendation results constrained by both the VA emotional path and the tag semantic path, thereby improving the emotional adaptability and semantic interpretability of the recommendation results.
[0019] 5. This invention constructs a VA-Tag joint sorting function. The algorithm integrates the target emotion proximity item, the current emotion association item, and the tag semantic matching item to calculate the comprehensive score of candidate music, and outputs a Top-K music recommendation list based on this score.
[0020] Compared to ranking based on single similarity or single tag, this invention can comprehensively measure whether candidate music is close to the target emotion, whether it connects with the current emotion, and whether it conforms to the target tag distribution, thereby improving the overall adaptability of the recommendation ranking.
[0021] 6. This invention can output recommendation explanation information, including the user's current emotional state, target emotional state, the music VA prediction attribute of the recommended music, the prediction probability or display tag of the dominant music tag, the distance between candidate music and the target emotional state, the distance between candidate music and the current emotional state, and tag semantic matching score information. Therefore, users can not only obtain a Top-K recommended music list, but also understand why the system recommends that music, improving the transparency and interpretability of the recommendation results. Attached Figure Description
[0022] Figure 1 This is a general flowchart of an embodiment of the present invention.
[0023] Figure 2 This is a schematic diagram of the system structure according to an embodiment of the present invention.
[0024] Figure 3 This is a schematic diagram of VA space mood regulation and music recommendation in an embodiment of the present invention. Detailed Implementation
[0025] The following will describe in more detail an intelligent music recommendation method for user emotion regulation according to the present invention, with reference to the accompanying diagrams, which illustrate preferred embodiments of the invention. It should be understood that those skilled in the art can modify the invention described herein while still achieving its advantageous effects. Therefore, the following description should be understood as being of general knowledge to those skilled in the art and is not intended to limit the invention.
[0026] Terminology definitions and symbol conventions: - User-input text refers to the natural language text that a user uses to describe their current mood or state. That is... Figures 1-3 "in ".
[0027] - Regulation strategies refer to the direction of emotion regulation goals chosen by the user, which include at least one of the following: maintain, calm, comfort, and energize.
[0028] - Adjustment intensity, used to control the degree to which the current emotional state moves toward the target emotional state, with a value range of [0,1].
[0029] - The candidate music set refers to the music library that the recommendation system needs to filter, containing at least one candidate music. .
[0030] in, , , , and These are the symbolic representations for the current emotional state, the strategy anchor, the target emotional state, the music VA prediction attribute, and the music Tag prediction attribute, respectively.
[0031] - Current emotional state, represented by a two-dimensional continuous coordinate system (VA) predicted by the system based on the user's input text. =( ); - Strategy anchor point, corresponding to the selected adjustment strategy. A reference coordinate point in VA space represents the ideal endpoint direction of emotion regulation under a given regulation strategy.
[0032] - Target emotional state, referring to the state achieved through the formula The calculated VA continuous coordinates represent the emotional position that the system hopes the user will ultimately reach; -Music VA prediction attribute, targeting candidate music The continuous coordinates of VA predicted by the audio regression model are represented as follows: =( This coordinate represents the actual location of the song within the VA space. Each candidate song has its own... . -Valence; - Wake-up rate.
[0033] The emotional space consists of VA (Valence) and Arousal.
[0034] - Music tag prediction attribute, i.e., pointer to candidate music This is a multidimensional probability vector predicted by a multi-label classification model. If the preset label system has... Each label, then =[ , ,..., ],in Represents music Belongs to the label The model predicts probabilities. This vector itself is not a coordinate and is not directly labeled in the VA space.
[0035] - Music tags, which are specific semantic categories within a pre-defined tagging system, for example... , , wait.
[0036] -Tag prototype, i.e., for music tags This refers to the coordinates of the tag within the VA space. These coordinates are not the position of a specific song, but rather the collective centroid of all songs predicted to be highly associated with that tag within the VA space. Each tag in each pre-defined tagging system... There is one and only one .
[0037] -Target Tag distribution, based on With each The probability distribution formed after normalizing the VA distance is expressed as: =[ , ,..., ,..., ].
[0038] - Joint sorting function, used for comprehensive evaluation of candidate music. The fitting score is calculated using a function that assigns a higher score, resulting in a higher recommendation priority.
[0039] Among them, regarding and The essential difference: and Although both are two-dimensional coordinates in VA space, they have fundamentally different meanings: It is an "individual" coordinate: belonging to a specific candidate piece of music. It's the emotional attribute of the music itself. There are tens of thousands of candidate songs, meaning there are tens of thousands of... .
[0040] It is the focus of the "group": belonging to a certain music tag. , is the average position of all songs predicted to be highly associated with this tag in the VA space.
[0041] Coordinate conventions in VA space: The VA space is a two-dimensional continuous plane. The horizontal axis represents valence, with the right end representing positive / pleasure and the left end representing negative / unpleasantness. The vertical axis represents arousal, with the upper end representing high arousal / excitement and the lower end representing low arousal / calmness.
[0042] like Figures 1-3 The intelligent music recommendation method based on user emotion regulation includes the following steps: I. Offline preprocessing stage (executed during system initialization or music library update, without relying on any user input): This stage involves the candidate music collection. Each of the candidate music Perform pre-calculation and store all calculation results in the database for querying and reference during the online recommendation phase.
[0043] (a) Pre-calculated music VA prediction attributes : Candidate music The audio signal is used to extract features, which are then fed into a pre-trained music VA prediction model (composed of an audio encoder and a continuous sentiment regression layer). The model outputs the two-dimensional coordinates of the music in the VA space. =( ).Will The song ID is associated with the song ID and stored in the database.
[0044] (ii) Pre-calculation of music tag prediction attributes : Candidate music The audio signal is used to extract features, which are then fed into a pre-trained music tag prediction model (multi-label classification model). The model outputs the probability distribution of the music on a pre-defined tag system. =[ , ,..., ].Will The song ID is associated with the song ID and stored in the database.
[0045] In this embodiment, the preset tag system includes 26 music tags, covering three semantic categories: Genre, Instrument, and Mood / Theme.
[0046] Description of the preset tag system: In one implementation, the music tag system includes three categories of tags: Genre, Instrument, and Mood / Theme.
[0047] The Genre category includes 10 tags: ambient, classical, electronic, folk, hip-hop, jazz, metal, pop, rock, and soundtrack; the Instrument category includes 7 tags: bass, drums, guitar, piano, strings, synthesizer, and voice; and the Mood / Theme category includes 9 tags: dark, dream, energetic, epic, film, happy, love, relaxing, and sad. These three categories together contain 26 tags.
[0048] Tag predicts attributes To predict probabilities for the model, candidate music does not require real human labels.
[0049] In this embodiment, the preset tag system uses the aforementioned 26 music tags. A 26-dimensional probability vector (III) Building a Tag Prototype : Based on the full set of candidate music obtained from the above pre-calculation... and For each music tag in the preset tag system Construct its mapping point in the VA space, i.e., the Tag prototype. .
[0050] For any music tag (For example ==" The specific construction method is as follows: 1. Traverse the candidate music set Take out each song from all the music in the tags. The predicted probability, i.e. The middle corresponds to the label The amount .
[0051] 2. According to Sort all candidate music from highest to lowest.
[0052] 3. Select the top 5% of candidate music tracks as the high-confidence sample set. .
[0053] 4. Each piece of music in it, with its As weights, predict its music VA attributes The weighted average is calculated using the following formula:
[0054] 5. Calculated two-dimensional coordinates That is, music tags The prototype of the tag.
[0055] After the calculation is completed, all Stored in a database.
[0056] In this way, the system can map discrete tag semantics to a continuous VA emotion space, establishing a mapping relationship between music semantic tags and the emotion space.
[0057] Figure 3 Correspondence: Figure 3 The names of the tags labeled within the VA space, including happy, classical, dream, sad, relaxing, ambient, piano, pop, rock, energetic, etc., are represented by the center of the circle to which they belong, which is the prototype of the corresponding tag. .
[0058] II. Online Recommendation Phase (Executed when responding to a single user request).
[0059] This stage is executed in real time when a user initiates a recommendation request, and reads the data stored in the offline preprocessing stage by querying the database.
[0060] Step 1: Obtain input parameters.
[0061] The system receives the following four inputs: User input text For example, "I'm so tired from work today, and I feel really down."
[0062] Regulation strategy In this embodiment, the value is energize (energize strategy).
[0063] Adjusting intensity In this embodiment, the value is 0.8.
[0064] Candidate Music Collection It includes all music to be recommended in the system's music library.
[0065] The above inputs correspond to Figure 1 The input module and Figure 2 The user interface module in [the system / platform].
[0066] Step 2: Estimate the current emotional state .
[0067] The system will input text from the user. The text is fed into a pre-trained text VA prediction model. This model consists of a text feature encoder and a continuous sentiment regression layer, which maps natural language text to the VA continuous sentiment space and outputs a two-dimensional coordinate as the current sentiment state. .
[0068] The current emotional state is a two-dimensional continuous emotional state consisting of valence and arousal components. The valence component represents the degree of positive or negative emotion, while the arousal component represents the degree of excitement or calmness.
[0069] In this embodiment, the user input text "So excited after the party, but need to calm down" is predicted by the model to obtain... =(-0.60, -0.5), which means it is in the lower left of the VA space, in the (low valence, low arousal) area, corresponding to the emotional state of "fatigue, depression, lack of vitality".
[0070] Figure 3 Correspondence: Figure 3 The solid circle marked "Current" has coordinates as follows: =(-0.60, -0.5), located in the lower left of the VA space (both the horizontal and vertical coordinates are negative).
[0071] Step 3: Generate the target emotional state .
[0072] This step first determines the strategy anchor point. Then, the target emotional state is generated through linear interpolation. .
[0073] (3.1) Determine the strategy anchor point : Adjustment strategies include maintaining the current state strategy. Calm Strategy comforting strategies and revitalization strategies At least one of them.
[0074] like = (Maintaining the current state), then = This keeps the target's emotional state close to its current state.
[0075] like for , or In one approach, the system retrieves the set of representative tags corresponding to a pre-defined strategy-tag mapping relationship, and reads the tag prototypes of these representative tags from the database. Then take its geometric center as .
[0076] The geometric center is the arithmetic mean of the coordinates of all the prototype coordinates of the tags, i.e. .
[0077] Preset labels for calming strategies include: ambient, relaxing, piano, dream, classical; Pre-defined labels for comfort strategies include piano, sad, classical, dream, and relaxing.
[0078] In this embodiment, = (Cheer-up strategy).
[0079] The default representative labels for boosting strategies include:
[0080] The system reads the tag prototypes of these five tags from the database (i.e., , , , , ), calculate their arithmetic mean, and obtain the policy anchor of the calm policy. =(0.7, 0.8), this point is located in the lower right of the VA space (high cost, low wake-up).
[0081] (3.2) Calculate the target emotional state : Using linear interpolation formula:
[0082] Substituting the value s=0.8 into this embodiment, =(-0.60, -0.5), =(0.7, 0.8): Valence component: (1-0.8)×(-0.6)+0.8×0.7=0.44; Arousal component: (1-0.8)×(-0.5)+0.8×0.80=0.54; Obtain the target emotional state =(0.44, 0.54).
[0083] Figure 3 Correspondence: Figure 3 The middle label is The coordinates of the hollow circle point are... Both the horizontal and vertical axes are positive, located in the upper right of the VA space (high valence, high arousal), achieving emotional regulation from "fatigue and depression" to "positive and uplifting".
[0084] Step 4: Obtain the Musical VA Prediction Attributes of Candidate Music : The system queries and reads each candidate music from the candidate music set M in the database. Pre-stored music VA prediction attributes =( ).
[0085] Each candidate music track output in this step This will be used in steps 7.1 (target emotion proximity item A(i)) and 7.2 (current emotion association item B(i)) to measure the relationship between the music and the target emotional state, respectively. The degree of closeness, and the relationship between the music and the current emotional state. The degree of connection.
[0086] It should be emphasized that m1, m2, m3, and m4 are only... Figure 3 The labels in the diagram do not represent any ranking. m1, m2, m3, and m4 are used to represent a subset of music in the candidate music set. In actual system operation, steps 4-7 perform calculations on all candidate music in M.
[0087] Step 5: Obtain the music tag prediction attribute of the candidate music : The system queries and reads the candidate music set from the database. Each of the candidate music Pre-stored music tag prediction attributes .
[0088] It is a 26-dimensional probability vector, that is, for each candidate music A probability value was predicted for all 26 tags.
[0089] Important Note: It is a multidimensional probability vector, and is not itself a coordinate in VA space, therefore it is not directly labeled. Figure 3 It is located in the VA space. It is only used as the computation input for the Tag semantic matching item in step 7.
[0090] Figure 3 Correspondence: Figure 3 Annotated in VA space For example, the coordinates of the music item's location are the coordinates of the corresponding music retrieved from the database. .
[0091] Each candidate music track output in this step This will be used in conjunction with the target tag distribution D output from step 6 in step 7.3 (tag semantic matching). ), calculated through dot product = * This measures the degree to which the music's tag attributes match the user's target semantic preferences.
[0092] Step 6: Generate the target tag distribution .
[0093] The system generates the target emotional state based on step 3. and the tag prototypes pre-stored in the database during the offline preprocessing stage Generate target tag distribution Specifically, the system targets each tag in the preset tagging system. Calculate their weights in the target tag distribution. The weights of all labels together constitute the complete distribution. .
[0094] The specific operating method is as follows: calculate With each Tag prototype Euclidean distance in VA space:
[0095] Distance Transformed into an activation value using a negative exponential mapping function:
[0096] This mapping guarantees distance. The closer the tag prototype, the greater its activation value.
[0097] 3. Perform Softmax normalization on the activation values of all tags to obtain the target tag distribution:
[0098] The weights of each label calculated above Combining these elements, we obtain the complete target tag distribution:
[0099] In this embodiment =(0.44, 0.54), located in the upper right corner of the VA space (high efficiency, high wake-up rate).
[0100] System Calculation In the VA space, the nearest label prototype to this point is... , .
[0101] Therefore, in the normalized target tag distribution D... The three tags received the highest weight.
[0102] by For example, its highest value indicates that the system semantically prefers the "energetic" music style; Secondly; the remaining 23 tags (such as The weight of low-wake-up tags is relatively low.
[0103] Figure 3 Correspondence: Figure 3 middle The nearest label name around the point (e.g.) The tag with the highest weight in the target tag distribution D is denoted as .
[0104] The target tag distribution output in this step This will be compared with each candidate music output in step 5. Used together in step 7.3 (Tag semantic matching items) ), calculated through dot product = * This measures the degree to which the music's tag attributes match the user's target semantic preferences.
[0105] Step 7: Calculate the joint sorting function : The system targets the candidate music set Each of the candidate music Calculate its comprehensive recommendation score The joint sorting function is a weighted combination of the following three factors: (7.1) Target Emotion Proximity Item : Measuring candidate music of With the target emotional state Proximity:
[0106] The closer the distance, The closer it is to 1.
[0107] (7.2) Current sentiment related items : Measuring candidate music of With current emotional state Degree of connection:
[0108] This ensures that the recommended music does not abruptly change with the user's current mood, guaranteeing a smooth transition in the auditory experience.
[0109] (7.3) Tag semantic matching items : Measuring candidate music of Distribution of target tags Degree of matching between:
[0110] This item is The dot product with D.
[0111] Indicates label In target tag distribution The weights in the equation.
[0112] Indicates candidate music Belongs to the label The predicted probability, i.e. The vector corresponding to the label The amount.
[0113] If a song is Labels with medium to high weights have a higher prediction probability, and therefore score higher for that item.
[0114] (7.4) Three-item weighted fusion: The above three items are merged by pre-setting weight parameters α, β, and γ:
[0115] in + + =1, ≥0, ≥0, ≥0.
[0116] In this embodiment, the following is taken =0.4, =0.2, =0.4.
[0117] Step 8: Output the Top-K music recommendation list: The system calculates based on step 7. For the candidate music set Sort all candidate music in descending order and output the K music with the highest scores as the recommendation list.
[0118] Figure 3 Correspondence: Figure 3 Each point within the VA space represents a candidate piece of music. After sorting in step 8, the music is closest to the target emotional state. The candidate music corresponding to that point is the overall score. The symbols (such as highlighted marks) represent the final output recommended music, and their coordinates represent the location of that recommended music. .
[0119] It should be noted that, Figure 3 The points marked m1, m2, m3, and m4 represent only a portion of the candidate music set and are not directly equivalent to recommended music. Whether m1~m4 ultimately become recommended music depends on each step in step 7. The calculation results and the sorting in step 8.
[0120] While outputting the Top-K music recommendation list, the system simultaneously outputs recommendation explanations, enabling users to understand why the system recommends these music tracks and improving the transparency and credibility of the recommendation system.
[0121] The recommended explanation information includes, but is not limited to, the following: Estimated current emotional state of the user The numerical value and its position in the VA space are described; Target emotional state generated The numerical value and its position in the VA space are described; Music VA prediction attribute of each recommended music The value; The dominant music tag for each recommended song (i.e., The 1-2 tags with the highest probability in the middle and their predicted probabilities; The VA distance between each recommended song and the target emotional state ; The VA distance between each recommended song and the current emotional state ; Target sentiment proximity term in the joint ranking function Current emotion related items semantic matching items with tags Each of them has a score.
[0122] In other embodiments: When a user enters text indicating "happy and energetic after the party" and selects the strategy of maintaining the current state, the system estimates the current emotional state based on the user's text, keeps the target emotional state near the current emotional state, and then recommends music that matches the current positive and highly aroused state. When a user enters text indicating "excitement and rapid heartbeat after receiving good news" and selects a calming strategy, the system generates a target emotional state with lower arousal and recommends music with display tags such as classical, piano, and film. When a user enters text indicating "tired, stressed" and selects a comfort strategy, the system generates a target emotional state that is closer to comfort and calm, and recommends music with display tags such as classical, piano, and relaxing. When a user enters the text "I'm in a good mood and hope to have more motivation" and selects an uplifting strategy, the system generates a target emotional state with a higher arousal orientation and recommends music with display tags such as pop, bass, and happy.
[0123] The above embodiments illustrate that the present invention can generate different target emotional states based on user text, adjustment strategies, and adjustment intensity, and output music recommendation results with differentiated explanatory information.
[0124] The above are merely preferred embodiments of the present invention and do not constitute any limitation on the present invention. Any equivalent substitutions or modifications made by those skilled in the art to the technical solutions and content disclosed in the present invention without departing from the scope of the present invention shall be deemed to have remained within the protection scope of the present invention.
Claims
1. A method for intelligent music recommendation oriented to user emotion regulation, characterized in that, Including the following steps: Step 1: Obtain user input text, adjustment strategy, adjustment intensity, and candidate music set; Step 2, estimating a current emotional state from the user input text , the current emotional state being a coordinate in a two-dimensional continuous emotional space of valence- arousal. Step 3: Based on the current emotional state Regulation strategies and regulation intensity generate target emotional states Among them, strategy anchor point Based on the Tag prototype corresponding to the preset representative label Confirmed, the Tag prototype These are fixed values pre-stored in the database; Step 4: Obtain the music VA prediction attribute for each candidate music in the candidate music set. The music VA prediction attribute The coordinates of the candidate music in the valence-arousal two-dimensional continuous emotion space; Step 5: Obtain the music tag prediction attribute for each candidate music in the candidate music set. The music tag prediction attribute The multidimensional probability vector is formed by the predicted probabilities of the candidate music on each tag in the preset tag system; Step 6, generating a target Tag profile based on the target emotional state and each of the pre-stored Tag prototypes generating a target Tag profile ; Step 7: Based on the current emotional state The target emotional state The music VA prediction attribute The music tag prediction attribute and the target tag distribution Calculate the joint ranking function for each candidate music. The joint sorting function This includes target emotion proximity items, current emotion association items, and tag semantic matching items; Step 8: According to the joint sorting function Sort the candidate music in the candidate music set and output a Top-K music recommendation list.
2. The intelligent music recommendation method for user emotion regulation according to claim 1, characterized in that, in step 3 according to the current emotional state , the adjustment strategy and the adjustment intensity to generate a target emotional state , comprising: determining a policy anchor point according to the regulation strategy , the target emotional state is calculated by formula , wherein, is the regulation intensity, and the value range is [0, 1]; The regulation strategy includes at least one of the following: maintaining the current state strategy, calming strategy, comforting strategy, and cheering strategy; When the adjustment strategy is to maintain the current state, the strategy anchor point When the regulation strategy is a calming strategy, a comforting strategy, or an uplifting strategy, the strategy anchor point... Tag prototypes representing the corresponding strategies Sure.
3. The intelligent music recommendation method for user emotion regulation according to claim 1, characterized in that, The Tag prototype Pre-constructed by: Step A, obtaining music VA prediction attributes of each candidate music in the candidate music set and music Tag prediction attributes ; Step B: For each type of label in the preset label system According to the candidate music in the tags On the predicted probability Sort the candidate music from highest to lowest, and select the top 5% as the high-confidence sample set. ; Step C: Using the predicted probability of the high-confidence sample As weights, the music VA prediction attributes of the high-confidence samples are... Perform a weighted average to obtain the label. Tag prototype .
4. The intelligent music recommendation method for user emotion regulation according to claim 3, characterized in that, The music VA prediction attribute and the music tag prediction attribute These are fixed values pre-calculated and stored in a database through audio feature extraction and model prediction; The online recommendation phase queries the database for reading the and the .
5. The intelligent music recommendation method for user emotion regulation according to claim 1, characterized in that, In step 6, based on the target emotional state and the pre-stored prototypes of each of the aforementioned tags Generate target tag distribution ,include: Step 6A: Calculate the target emotional state Tag prototypes for each type of label Euclidean distance in valence-arousal space; Step 6B: Convert the Euclidean distance into an activation value using a negative exponential mapping function; Step 6C: Perform Softmax normalization on the activation values of all tags to obtain the target tag distribution.
6. The intelligent music recommendation method for user emotion regulation according to claim 1, characterized in that, The joint sorting function described in step 7 include: Target Emotion Proximity Item The music VA prediction attribute used to measure the candidate music. With the target emotional state The degree of closeness; Current sentiment related items The music VA prediction attribute used to measure the candidate music. With the current emotional state The degree of connection; Tag semantic matching items The music tag prediction attribute used to measure the candidate music. Distribution of the target tag The degree of matching; The joint sorting function is: ,in , , For preset weight parameters, and .
7. An intelligent music recommendation system oriented towards user emotion regulation, characterized in that, include: The user interface module is used to obtain user input text, adjustment strategy, adjustment intensity, and candidate music set; The text sentiment estimation module is used to estimate the current sentiment state based on the user's input text. The current emotional state The coordinates are in the two-dimensional continuous emotion space of valence-arousal. The target emotion state generation module is used to generate emotions based on the current emotion state. The regulation strategy and the regulation intensity generate the target emotional state. ; A database is used to pre-store the music VA prediction attributes of each candidate music in the candidate music set. Music Tag Prediction Attributes And the tag prototype for each type of tag in the preset tag system. ; The audio feature extraction module is used to extract audio features from candidate music tracks; The music VA prediction module is used to output the music VA prediction attribute of the candidate music based on the audio features. ; The music tag prediction module is used to output the music tag prediction attribute of the candidate music based on the audio features. ; Tag prototype building module, used to predict attributes based on the music VA. and the music tag prediction attribute For each type of tag in the preset tag system, construct a tag prototype. ; The target tag distribution generation module is used to generate tags based on the target emotional state. and the pre-stored prototypes of each of the aforementioned tags Generate target tag distribution ; The joint sorting module is used to sort according to the current emotional state. The target emotional state The music VA prediction attribute The music tag prediction attribute and the target tag distribution Calculate the joint ranking function for each candidate music. and according to the joint sorting function Sort the candidate music tracks and output a Top-K music recommendation list; And a recommendation explanation information output module, used to output recommendation explanation information.
