Method, device and terminal for modeling music recommendation system based on dynamic user preferences

Abstract

The invention discloses a method, a device and a terminal for modeling a music recommendation system based on dynamic user preferences, and relates to the technical field of data processing. Comprising the following steps: collecting data through dialogue interaction and behavior analysis, and constructing an initial user preference model based on source weight, data freshness and behavior consistency; calculating the matching degree between the declaration preference and the actual song listening behavior data; if the matching degree is lower than the threshold value and meets the time condition, active dialogue interaction is triggered to obtain real-time intention feedback; performing semantic analysis on the feedback, correcting the initial user preference model according to an identification result, and updating the initial user preference model to a database; and generating recommended content according to the corrected model and scene. According to the method, the explicit dialogue data and the implicit behavior data of the user are acquired through dialogue interaction and behavior analysis dual modes, the credibility of different data sources is distinguished by using source weights, the data is filtered in combination with data freshness and behavior consistency, and the initial user preference model capable of objectively reflecting real demands of the user is constructed.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method, apparatus and terminal for a dynamic user preference modeling music recommendation system. Background Technology

[0002] With the popularization of digital music media technology, music recommendation systems have become an important way for users to obtain music content. Traditional music recommendation systems typically use collaborative filtering algorithms or content-based recommendation algorithms to build user models by analyzing implicit data such as users' historical listening records, click behavior, or favorites lists, or to generate static user profiles based on fixed interest tags selected by users during registration. Based on this, they predict the tracks that users may be interested in and generate a recommendation list.

[0003] However, the aforementioned recommendation methods based on a single data source or static model have limitations in practical applications. On the one hand, relying solely on historical implicit behavioral data makes it difficult to accurately distinguish between accidental clicks and genuine preferences, leading to a one-sided understanding of user intent. Furthermore, the lack of data accumulation in the initial stages of new user access results in low recommendation accuracy. On the other hand, users' music preferences are dynamic and change over time, depending on mood or environment. Existing static user profiles lack real-time deviation detection and correction mechanisms, failing to promptly perceive the differences between current user behavior and historical preference models, causing recommended content to lag behind the user's latest interests.

[0004] Furthermore, existing systems typically perform matching based on predefined, fixed sets of labels, lacking the ability to understand personalized, non-standardized, or novel style descriptions submitted by users, thus limiting the coverage and flexibility of recommendation systems. Therefore, how to construct a dynamic music recommendation mechanism that can integrate multimodal data and perceive and correct preference biases in real time is a problem that needs to be solved in this field. Summary of the Invention

[0005] The purpose of this invention is to provide a method, apparatus, and terminal for a dynamic user preference modeling music recommendation system, which at least solves one of the technical problems of lagging user preference model updates, low cold start recommendation accuracy, and inability to perceive the deviation between user preferences and actual behavior in real time.

[0006] To address the above problems, the present invention provides the following technical solution:

[0007] The first aspect of this invention provides a method for a dynamic user preference modeling music recommendation system, employing the following technical solution:

[0008] A method for modeling dynamic user preferences in a music recommendation system, applied to a music recommendation system including a music recommendation device and a user terminal, includes:

[0009] S10. Collect explicit dialogue data and implicit behavioral data of users through a dual-modal approach of dialogue interaction and behavior analysis, and construct an initial user preference model based on source weight, data freshness, and behavioral consistency.

[0010] S20. Obtain the user's actual listening behavior data and calculate the matching degree between the user's declared preferences in the initial user preference model and the actual listening behavior data;

[0011] S30. Detect whether the matching degree is lower than the matching degree threshold and meets the trigger time condition; if so, trigger active dialogue interaction and obtain the user's real-time intention feedback through voice inquiry.

[0012] S40. Perform semantic analysis on the real-time intent feedback, identify emotional features or new label attributes, revise the initial user preference model based on the identification results and update it to the unstructured preference database to obtain the revised user preference model.

[0013] S50: In response to the user's playback request, generate a music playlist or recommended content based on the revised user preference model and the current context.

[0014] By adopting the above technical solution, this invention constructs a recommendation system with feedback and update functions. Through the bimodal acquisition in step S10, the system combines the user's explicitly expressed subjective intentions with implicitly expressed objective behaviors to establish a multi-dimensional user preference model benchmark. Steps S20 to S40 introduce an active feedback mechanism based on deviation detection. By calculating the matching degree between statements and behaviors in real time, the system can identify when user preferences change or the current model fails, and immediately correct the model through active dialogue interaction, rather than waiting for the accumulation of behavioral data. This mechanism solves the problem that static models cannot adapt to changes in user interests, ensuring that recommended content remains consistent with the user's current true intentions and improving the adaptability of the recommendation system.

[0015] Furthermore, in step S10, the construction of the initial user preference model based on source weight, data freshness, and behavioral consistency includes: extracting preference tags from explicit dialogue data and implicit behavioral data, and assigning source weights to the preference tags, wherein the source weight value of preference tags from explicit dialogue data is set to be greater than the source weight value of preference tags from implicit behavioral data; calculating data freshness based on the timestamp of preference tag generation, and calculating behavioral consistency based on the recurrence frequency of preference tags in historical behavioral sequences; calculating the final confidence level by calculating the source weight, data freshness, and behavioral consistency, and writing preference tags with a final confidence level greater than the confidence level entry threshold into the initial user preference model.

[0016] By adopting the above technical solutions, the system establishes a multi-dimensional data evaluation system. By assigning higher weight to explicit dialogue, the system prioritizes the use of users' subjective expressions, reducing the probability of recommendation bias caused by accidental clicks. Introducing a data freshness parameter enables the model to have a time decay strategy, which can reduce the influence of outdated preferences; introducing behavioral consistency verification eliminates random noise. This comprehensive calculation logic ensures that the data written into the initial model has high credibility, providing a data foundation for recommendation calculation.

[0017] Furthermore, in step S20, calculating the matching degree between user-declared preferences in the initial user preference model and actual music listening behavior data includes: calculating the overlap ratio between the set of declared preferences in the initial user preference model and the set of actual behavior tags composed of actual music listening behavior data to obtain the tag overlap degree; detecting whether the style tag of the currently played song in the actual music listening behavior data is consistent with the style tag associated with the current context scene in the history record whose recurrence frequency exceeds the frequency threshold to obtain the scene matching degree; and generating the matching degree based on the weighted result of the tag overlap degree and the scene matching degree.

[0018] By adopting the above technical solutions, the system quantifies the consistency of users' words and actions from two dimensions: content and environment. Tag overlap reflects the stability of users' long-term preferences, while scenario matching reflects whether users' immediate needs are met in a specific context. The matching index generated through weighted calculation can determine whether recommended content matches users' preferences and current scenario needs, thus providing a quantitative basis for the system to determine whether to trigger proactive intervention.

[0019] Furthermore, in step S30, the triggering time condition is that the time interval since the last model update exceeds the minimum update time interval; the triggering of active dialogue interaction includes: when the matching degree is detected to be lower than the matching degree threshold and the triggering time condition is met, generating voice query content containing information of the current context scene.

[0020] By adopting the above technical solutions, the system introduces frequency control logic and scene awareness logic into the proactive interaction mechanism. Setting a minimum update interval avoids degrading the user experience due to frequent questioning. Generating question content containing scene information makes the system's questions more targeted. This contextualized interaction method can guide users to accurately express their current needs, thereby improving the effectiveness of feedback data.

[0021] Furthermore, in step S40, the step of performing semantic analysis on the real-time intent feedback, identifying emotional features or new label attributes, and correcting the initial user preference model and updating it to the unstructured preference database based on the identification results includes: extracting emotional state slots and demand slots from the real-time intent feedback using a natural language processing model; establishing a correlation between the attributes corresponding to the demand slots and the current context scene, and writing them into the unstructured preference database.

[0022] By adopting the above technical solution, the system achieves structured processing and storage of unstructured voice feedback. Through slot extraction technology, the system captures the implicit emotional features and specific needs in the user's language and associates these needs with specific contextual scenarios. This processing method enables the preference database to form a mapping relationship between scenarios and needs, allowing the system to invoke corresponding preference settings when encountering similar scenarios in the future, thus achieving scenario-based recommendations.

[0023] Furthermore, the unstructured preference database adopts a key-value pair hierarchical structure; in step S40, the identification of sentiment features or new tag attributes further includes: when the identified new tag attribute is not included in the local standard tag set, marking the new tag attribute as an unknown tag, and querying an external knowledge base to obtain the description information of the unknown tag; extracting acoustic feature attribute words from the description information, mapping the unknown tag to a locally executable combined tag, and writing the mapping relationship between the unknown tag and the combined tag into the unstructured preference database.

[0024] By adopting the above technical solution, the system solves the problem of new tag recognition. When users use style terms not defined by the system, the system utilizes an external knowledge base for semantic mapping, transforming unknown concepts into acoustic feature combinations that the system can understand. This gives the system adaptive learning capabilities and scalability, allowing it to respond to users' evolving language habits and musical needs without frequent manual updates to the tag library.

[0025] Furthermore, in step S50, responding to the user's playback request and generating a music playlist or recommended content based on the modified user preference model and the current context includes: identifying the intent type of the user's playback request; if the intent type is a general request scenario, then retrieving music style tags associated with the current context from the unstructured preference database and filtering tracks that match the music style tags; if the intent type is a new song exploration scenario, then calculating the similarity between the acoustic feature vectors of songs whose historical playback frequency exceeds a frequency threshold and the acoustic feature vectors of candidate songs based on the acoustic feature vectors of those songs and filtering tracks with a similarity greater than a filtering threshold; if the intent type is a specified type scenario, then retrieving the specified main tag in the hierarchical structure of the unstructured preference database and removing sub-genre tracks with negative feedback records.

[0026] By adopting the above technical solutions, the system constructs a differentiated recommendation strategy. For ambiguous commands, it uses scene association logic for recommendations; for exploratory commands, it uses acoustic feature similarity to recommend potentially liked new content; and for explicit commands, it uses the hierarchical structure of the database for negative filtering. This processing logic covers various user scenarios and improves the system's recommendation accuracy.

[0027] Furthermore, the unstructured preference database includes a genre field, an artist field, and a scene field; the step of removing tracks from sub-genre fields with negative feedback records includes: checking the sub-genre exclusion list under the genre field and filtering out tracks that fall into the sub-genre exclusion list.

[0028] By adopting the above technical solution, the system achieves fine-grained filtering management under categories. This solution allows users to retain a mainstream preference while excluding specific subgenres within it. Through filtering the subgenre exclusion list, the system can generate playlists that meet the user's specific requirements.

[0029] A second aspect of the present invention provides an apparatus for a dynamic user preference modeling music recommendation system, which adopts the following technical solution:

[0030] An apparatus for a dynamic user preference modeling music recommendation system, comprising:

[0031] The preference collection module is configured to collect explicit dialogue data and implicit behavioral data of users through a dual-modal approach of dialogue interaction and behavior analysis, and to build an initial user preference model based on source weight, data freshness, and behavioral consistency.

[0032] The matching calculation engine is configured to obtain the user's actual music listening behavior data and calculate the matching degree between the user's declared preferences in the initial user preference model and the actual music listening behavior data;

[0033] The negative feedback control module is configured to detect whether the matching degree is lower than the matching degree threshold and the trigger time condition is met; if so, it triggers active dialogue interaction, obtains the user's real-time intent feedback through voice inquiry, performs semantic analysis on the real-time intent feedback, identifies emotional features or new label attributes, and corrects the initial user preference model based on the recognition results to obtain the corrected user preference model.

[0034] An unstructured preference database, configured to store the initial user preference model and the revised user preference model;

[0035] The music execution module is configured to respond to user playback requests by generating music playlists or recommended content based on the revised user preference model and the current context.

[0036] By adopting the above technical solution, the device realizes the preference modeling logic through modular collaboration. The preference acquisition module and matching calculation engine are responsible for data input and monitoring; the negative feedback control module is responsible for system adjustment; the unstructured preference database is responsible for data storage and reading / writing; and the music execution module is responsible for service output. The cooperation of these modules ensures the efficiency and stability of the system when processing user data, providing intelligent support for terminal devices.

[0037] A third aspect of this invention provides a terminal for a dynamic user preference modeling music recommendation system, employing the following technical solution:

[0038] A terminal for a dynamic user preference modeling music recommendation system includes:

[0039] The audio acquisition unit is configured to collect explicit dialogue data and real-time intent feedback from users.

[0040] The audio output unit is configured to play music playlists or recommended content, as well as play voice-activated questions during proactive dialogue interactions.

[0041] The positioning and sensor unit is configured to collect sensor data.

[0042] A communication unit configured to transmit data with a music recommendation device;

[0043] The processor is configured to generate the current context scene based on sensor data; and is configured to respond to the active interaction command issued by the music recommendation device, control the audio output unit to play the voice inquiry content, and control the audio acquisition unit to obtain real-time intent feedback within a set sampling time window.

[0044] By adopting the above technical solution, this terminal, serving as the carrier for user interaction with the recommendation system, leverages the perception and interaction capabilities of the edge device. The positioning and sensor units provide environmental context information; processor control ensures the timing of proactive interaction, and by controlling the audio acquisition time window, it guarantees the collection of feedback data and optimizes device power management. This edge-cloud collaborative architecture leverages the hardware advantages of the terminal and the computing power advantages of the cloud.

[0045] The above solution achieves the following beneficial technical effects:

[0046] This application collects explicit dialogue data and implicit behavioral data from users through a dual-modal approach of dialogue interaction and behavioral analysis. It constructs an initial user preference model based on source weight, data freshness, and behavioral consistency. The source weight is used to distinguish the credibility of different data sources, and data freshness and behavioral consistency are combined to filter out outdated or incidental behavioral data, thereby constructing an initial user preference model that can objectively reflect the user's real needs.

[0047] This application calculates the matching degree between user-declared preferences in the initial user preference model and actual music listening behavior data. When the matching degree is lower than the matching degree threshold and the triggering time condition is met, an active dialogue interaction is triggered to obtain real-time user intent feedback to correct the initial user preference model. This establishes a model update mechanism based on deviation detection. When the user's actual behavior deviates from the existing model, the latest user intent is captured through voice inquiry, solving the problem that static models cannot adapt to dynamic changes in user preferences.

[0048] This application maps unknown tags not included in the local standard tag set to locally executable combined tags by querying an external knowledge base, and writes the mapping relationship into an unstructured preference database with a key-value pair hierarchy. By using the external knowledge base, the tag system is extended, enabling the system to understand and process new tag attributes or non-standardized descriptions used by users. This solves the problem that a fixed tag set cannot cover the diverse expressions of users and improves the system's ability to understand users' personalized preferences. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating a dynamic user preference modeling music recommendation method according to an embodiment of the present invention.

[0050] Figure 2 This is a schematic diagram of the preference confidence modeling process according to an embodiment of the present invention.

[0051] Figure 3 This is a schematic diagram of the dynamic matching degree calculation process according to an embodiment of the present invention.

[0052] Figure 4 This is a schematic diagram of the active interaction and intent correction process according to an embodiment of the present invention.

[0053] Figure 5 This is a schematic diagram of an embodiment of the unstructured data storage and tag expansion process of the present invention.

[0054] Figure 6 This is a schematic diagram of the scenario-based music recommendation execution strategy flow according to an embodiment of the present invention.

[0055] Figure 7 This is a structural block diagram of an apparatus for a dynamic user preference modeling music recommendation system according to an embodiment of the present invention.

[0056] Figure 8 This is a schematic diagram of the terminal structure of a dynamic user preference modeling music recommendation system according to an embodiment of the present invention. Detailed Implementation

[0057] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0058] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0059] This invention applies to a music recommendation system architecture based on dynamic user preference modeling. This architecture includes a music recommendation device and a user terminal. The music recommendation device is deployed on a network-side server or cloud computing platform, and the user terminal is an electronic device directly operated by the user. The music recommendation device and the user terminal are connected via a communication network, including a mobile communication network, a wireless local area network, or a wired network, for data transmission between them.

[0060] The user terminal, acting as a front-end device for data acquisition and interaction, is equipped with an audio acquisition unit, an audio output unit, and a sensor array. The user terminal collects multimodal data from the user, including voice interaction data, data on media content playback control behavior, and sensor data reflecting the current environmental state. The user terminal then uploads the collected data to the music recommendation device via a communication network.

[0061] The music recommendation device, acting as a data processing and decision-making center, is equipped with memory and a processor. It receives data uploaded from user terminals and constructs a bimodal user preference model based on dialogue content and behavioral data. The device periodically calculates the degree of matching between users' declared preferences and actual behavioral data.

[0062] When the matching degree calculated by the music recommendation device is lower than a preset threshold and the negative feedback trigger condition is met, the music recommendation device generates an active interaction command. This active interaction command is sent to the user terminal. In response to the command, the user terminal plays the voice query content through the audio output unit and obtains the user's voice response through the audio acquisition unit.

[0063] The music recommendation system receives and analyzes the user's voice responses, extracting emotional features and intent information to update the user preference model. Based on the updated user preference model and the current user scenario, the system generates a music recommendation list or playback command, which is then sent to the user's terminal for playback.

[0064] In a practical application environment, the user terminal is an in-vehicle terminal, smartphone, or smart speaker. The user terminal uses its built-in GPS, accelerometer, and system clock to generate contextual scene labels. For example, when the user terminal detects that it is currently in a commuting scenario, it records the user's song selection and skipping behavior in real time. If the music recommendation device detects a discrepancy between the user's actual listening style in the commuting scenario and the preference labels stored in the database, the system automatically triggers a feedback update process. Through voice interaction, it confirms the user's current actual needs, thereby correcting the preference model and adjusting subsequent recommendation strategies.

[0065] See attached document Figure 1 This invention provides a dynamic user preference modeling music recommendation method, which is applied to a music recommendation system based on dynamic user preference modeling, and includes the following steps:

[0066] S10. Collect user preference data through a dual-modal approach of dialogue interaction and behavior analysis, and construct an initial user preference model based on source weight, data freshness, and behavioral consistency.

[0067] S20. Obtain the user's actual listening behavior data and calculate the matching degree between the user's declared preferences and the actual listening behavior data;

[0068] S30. Detect whether the matching degree is lower than the preset threshold and meets the trigger time condition; if so, trigger active dialogue interaction and obtain the user's real-time intention feedback through voice inquiry.

[0069] S40. Perform semantic analysis on real-time intent feedback, identify emotional features or new label attributes, correct the user preference model based on the identification results, and update it to the unstructured preference database.

[0070] S50: In response to the user's playback request, generate a music playlist or recommended content based on the revised user preference model and the current context.

[0071] The specific implementation principles and technical details of each of the above steps will be explained in detail below.

[0072] See attached document Figure 2 To address the bias in preference models caused by a single data source, this embodiment employs a dual-modal data acquisition strategy and introduces a confidence calculation model to weight preference labels. The specific execution process includes the following sub-steps:

[0073] S101: Collect the user's explicit dialogue data and implicit behavioral data, and extract preference tags respectively. The system receives the user's voice commands and dialogue content through the terminal's audio acquisition unit, converts the voice signal into a text sequence using automatic speech recognition technology, and then extracts explicit preference tags through natural language processing. For the specific implementation of automatic speech recognition and basic natural language processing, those skilled in the art can use existing deep learning models, which will not be elaborated here. Simultaneously, the system obtains the user's historical playback records, skipped records, and favorite records through a behavior record database, and extracts implicit preference tags from them. Explicit dialogue data corresponds to the user's subjective intentions, while implicit behavioral data corresponds to the user's actual operating habits.

[0074] S102, assign source weights to the extracted preference tags based on the data source type. In this embodiment, the credibility of explicit dialogue data is set higher than that of implicit behavioral data. For preference tags originating from user-initiated voice dialogue, the system sets a first source weight coefficient; for preference tags originating from background playback behavior, the system sets a second source weight coefficient, wherein the first source weight coefficient is greater than the second source weight coefficient. Specifically, the dialogue source weight is set to 0.7, and the behavioral source weight is set to 0.3, so that preferences confirmed by voice have a higher weight influence in the model calculation.

[0075] S103, calculate the data freshness and behavioral consistency of preference tags. Data freshness is used to quantify the impact of the time dimension on the validity of preferences. The system obtains the timestamp of the preference tag's generation. Data generated within a preset recent time period (e.g., the last 7 days) is assigned a higher freshness score; data generated in a more distant time period (e.g., 30 days ago) has its freshness score reduced by a preset decay ratio. In this embodiment, the weight of data from the last 7 days is set to twice the weight of data from 30 days ago. Behavioral consistency is used to quantify the stability of the same tag appearing in different time periods, obtained by calculating the recurrence frequency of the tag in historical behavioral sequences.

[0076] S104, based on source weight, data freshness, and behavioral consistency, uses a preference confidence calculation model to generate the final preference confidence score. This confidence score is used to measure whether a specific preference tag should be included in the current user profile. The preference confidence calculation model is defined as follows:

[0077] ;

[0078] The meanings of the symbols in the formula are as follows:

[0079] This indicates the final confidence level of the preference label;

[0080] Indicates the source weight, which is determined based on the data source type;

[0081] Indicates data freshness, determined according to the time decay rule;

[0082] Indicates behavioral consistency, reflecting the degree to which the label consistently matches the user's historical behavior;

[0083] , , These are the normalized weighting coefficients for each item, used to adjust the proportion of the influence of source, time, and consistency in the model.

[0084] S105, calculate the confidence level Compare with the preset entry threshold. If the confidence level of a certain preference label... If the value is greater than the preset entry threshold, the tag and its confidence score are written into the unstructured preference database; if the value is less than the threshold, it is marked as pending verification or excluded, thereby constructing an initial user preference model that includes the confidence score dimension.

[0085] See attached document Figure 3 To address the issue of insufficient timeliness in updating existing user profiles, this embodiment periodically performs consistency checks through a matching calculation engine to quantify the degree of fit between user-stated preferences and actual music listening behavior. The specific execution process includes the following sub-steps:

[0086] S201, obtain the user's declared preference set and actual behavior tag set within a preset time window. The system reads the user's currently active declared preference tag set from the preference database, denoted as... At the same time, extract historical interaction logs from the behavior record library showing that the user's actual playback frequency exceeded a preset threshold in the most recent period (such as the past 3 days or 1 week).

[0087] The data structure of this log is shown in Table 1:

[0088] Table 1. Structure Definition of Music Behavior Record Form

[0089] field name type describe record_id BIGINT Auto-incrementing primary key user_id VARCHAR(36) User ID song_id VARCHAR(32) Song unique code play_context JSON Playback scene {time period, location, emotion, etc.} tags JSON The system labels artists with tags such as [singer, genre, style]. timestamp TIMESTAMP Playback time

[0090] In this structure:

[0091] The play_context field stores a snapshot of the environment during playback in JSON format (e.g., {"Time":"Morning","Loc":"Car", "Mood":"Rush"}). This field provides direct data support for calculating scene matching degree (SctxSctx) in step S203.

[0092] The `tags` field stores the metadata tags for this song. The system counts the frequency of these tags within the time window defined by the `timestamp`, filters out tags whose frequency exceeds a preset threshold, and thus forms the actual behavior tag set. This is used for calculating the label overlap in step S202.

[0093] S202 calculates label overlap, used to quantify the proportion of actual behavior that covers declared preferences. Label overlap. Defined as the ratio of the number of elements in the intersection of the declared preference set and the actual behavior label set to the number of elements in the declared preference set. The calculation formula is as follows:

[0094] ;

[0095] In the formula, The cardinality of a set (i.e., the number of elements in the set). This represents the intersection operation of sets. This metric reflects the proportion of user-reported preferences that have been validated in recent behavior.

[0096] S203, Calculate the scene matching score, which is used to evaluate the degree of conformity between the current playback style and the historical settings of a specific scene. The system reads the context scene tag of the current time period (e.g., "commuting") and retrieves the high-frequency style tags associated with this scene in the historical records. If the style tag of the currently playing song matches the historical high-frequency tags of this scene, the scene matching score is calculated. Take the first value (e.g., 1.0); if inconsistent or conflicting, take the second value (e.g., 0.0). This step is used to identify whether the user's preferences have changed in a specific scenario.

[0097] S204, based on label overlap and scene matching, uses a weighted summation algorithm to generate the final matching score. Matching Score It is the core basis for triggering subsequent proactive interaction logic, and its calculation formula is as follows:

[0098] ;

[0099] The meanings of the symbols in the formula are as follows:

[0100] This represents a comprehensive matching score between user preferences and behaviors. This represents the label overlap calculated in S202; This represents the scene matching degree calculated in S203; Weighting coefficients representing the degree of label overlap; The weighting coefficients represent the scene matching degree.

[0101] In this embodiment, the following settings are provided: =0.7, =0.3, meaning the matching score primarily depends on the overlap of tags, while also considering scene adaptability. The system will calculate the matching score. Cache it in memory for subsequent comparison and decision with a preset threshold.

[0102] See attached document Figure 4 To address the lag and misjudgment issues inherent in relying solely on implicit behavior inference, this embodiment constructs an active interaction mechanism based on a negative feedback loop. When the mechanism detects that the match between a user's declared preferences and actual behavior is below standard, it proactively initiates a voice dialogue to correct the model. The specific execution process includes the following sub-steps:

[0103] S301, execute negative feedback to trigger logical decision. The system reads the matching score calculated in S204 in real time. and the last update time of the preference model To avoid frequently disturbing users, the system has set a matching threshold. and minimum update interval The system generates an active interaction trigger signal when the following logical expression is satisfied:

[0104] ;

[0105] In the formula, This represents the logical AND operation; Set a preset matching threshold for the system, for example, a value of 0.6; This is a preset time threshold, such as 3 days. This logic means that the active query process will only be initiated when the matching degree is below the threshold and a certain time has passed since the last update.

[0106] S302, generate a contextualized voice inquiry command and execute playback. In response to an active interaction trigger signal, the system generates an inquiry text based on the current playback context (e.g., "We've detected that you've been listening to classical music frequently lately; would you like us to adjust your commute playlist?"). This text is synthesized into a speech stream using text-to-speech technology and played through the terminal's audio output unit. To minimize user interference, the system monitors the audio playback status and controls the timing of the voice inquiry playback to coincide with the transition between two songs or when the user pauses playback.

[0107] S303: Collects the user's voice response and converts it into a text sequence. The system activates the terminal's audio acquisition unit to receive the user's voice feedback (e.g., "I've been under a lot of work pressure lately, I'd like to listen to something relaxing"). The automatic speech recognition module extracts and decodes the acoustic features of the collected voice data, outputting the corresponding natural language text sequence.

[0108] S304 performs deep semantic understanding and sentiment slot extraction on the text sequence. This step utilizes a pre-trained natural language processing model (such as BERT or Bi-LSTM networks) to process the text sequence output from S303. Specifically, the system maps the text sequence to a high-dimensional vector space for feature extraction, identifying sentiment features and intent slots. For example, it identifies the sentiment state slot {Sentiment:Negative / Stress} from the text "work pressure is high," and the demand slot {Requirement:Soothing} from "want to hear something relaxing." For the specific network structure and training process of the BERT or Bi-LSTM model, those skilled in the art can refer to existing deep learning frameworks for implementation, and will not be elaborated here.

[0109] S305, Correct the user preference model based on semantic analysis results. The system dynamically adjusts the mapping relationship in the preference database based on the extracted sentiment state and demand slots. Specifically, it establishes a direct association between the current scene tag (e.g., "commuting") and the identified demand attribute (e.g., "soothing / light music"), and marks the data source of this update as "dialogue interaction". According to the rules in step S102 above, the source weight of the new preference tag is updated to a high weight value, thereby updating the model weights and completing the correction process of the negative feedback loop.

[0110] See attached document Figure 5 To address the diversity of user preference descriptions and the continuous emergence of new music genres, this embodiment abandons the traditional fixed-field relational database and adopts an unstructured storage scheme based on JSON objects. It also combines this with an external knowledge base to dynamically expand tags. The specific execution process includes the following sub-steps:

[0111] S401, Construct an unstructured preference database based on JSON format. The system allocates an independent storage record for each user. To balance query efficiency and data flexibility, this embodiment adopts a hybrid storage mode, that is, embedding JSON data type in a relational table structure. Specifically, the table structure design of the user preference main table is shown in the following table:

[0112] Table 2. User Preferences Main Table Structure Definition Table

[0113] field name type describe Example user_id VARCHAR(36) User unique identifier u_123456789 pref_type ENUM Preference type GENRE pref_value JSON Preference value {"rock":["heavy_metal","folk_rock"]} confidence FLOAT Confidence level (0-1) 0.85 source ENUM source DIALOG last_update TIMESTAMP Last updated 2025-07-2114:30:00

[0114] In this structure:

[0115] The `pref_value` field employs a multi-level key-value pair structure to support fine-grained preference descriptions. To more clearly illustrate the hierarchical structure of this embodiment, a specific JSON data example is provided below:

[0116] {

[0117] "genre":{

[0118] "rock":["heavy_metal", "punk"], / / Mainstream rock, with subgenres favoring heavy metal and punk.

[0119] "pop":["k-pop", "j-pop"] / / Mainstream opinion favors K-pop and Japanese / Korean pop.

[0120] },

[0121] "artist":["Band A", "Singer B"], / / Specific artist preference

[0122] "scene":{

[0123] "commute":["energetic"], / / The "energetic" tag is associated with commuting scenarios.

[0124] "workout":["fast_tempo"] / / The "fast-paced" tag is associated with the motion scene.

[0125] }

[0126] }

[0127] With the above structure, when the system makes recommendations, it can not only match genres, but also perform secondary filtering based on related tags in the context, thereby achieving context-aware and accurate recommendations.

[0128] The confidence field stores the final confidence level calculated in step S104 above. This is used as the sorting criterion for subsequent recommendations.

[0129] The `source` field indicates whether the preference originates from "dialog" or "analysis," which is used to assign different weighting coefficients during updates. .

[0130] The last_update field is used for negative feedback logic decisions in step S301. The system calculates the time difference between the current time and this field to determine whether the minimum update interval is met.

[0131] S402, perform new tag recognition and local library matching detection. When the system extracts the user's intent keywords through the aforementioned natural language processing module, it compares the keywords with the existing standard tag set in the local database. If the keyword exists in the standard set, it is directly indexed; if the keyword (e.g., "metaverse rock" mentioned by the user) does not exist in the standard set, the system determines it as an "unknown tag" and triggers the tag expansion process.

[0132] S403 performs semantic queries and attribute associations on unknown tags based on an external knowledge base. The system connects to a pre-defined external general knowledge graph or open-domain encyclopedia database via an application programming interface. The unknown tags identified in S402 are used as query entities to obtain their descriptive information and associated nodes in the knowledge graph. For example, for "metacosmic rock," the system retrieves the definition, category, and related attributes of this term.

[0133] S404 performs feature reasoning and mapping on the query results. The system analyzes the descriptive information obtained in S403 and extracts attribute words related to acoustic features. Specifically, the system uses a semantic similarity algorithm to calculate the distance between the unknown tag and each tag in the local standard tag library. Taking "metacosmic rock" as an example, the system maps it to locally executable combined tags based on descriptions such as "virtual reality," "electronic music style," and "rock rhythm" from the knowledge base, such as {Primary_Style:"Rock",Sub_Style:"Electronic",Mood:"Futuristic"}. This process transforms user-defined non-standard vocabulary into music feature parameters that the system can understand.

[0134] In step S405, the expanded tag mapping relationship is dynamically written to the preference database. The system uses "unknown tag" as the new key and the inferred "mapping attribute" as the value to update the user's personalized dictionary. When the user uses the word again, the system directly calls the locally stored mapping relationship for recommendations, without repeatedly querying the external knowledge base, thus achieving adaptive expansion of preference tags and data accumulation.

[0135] See attached document Figure 6 To ensure that the final music content delivered not only meets the user's immediate needs but also aligns with their personalized aesthetic preferences in the long term, this embodiment executes differentiated filtering and sorting logic based on the user's request type. The specific execution process includes the following sub-steps:

[0136] S501 receives the user's playback request and parses the intent type. The system categorizes the user's voice commands or operation commands into three types: general request scenarios, new song exploration scenarios, and specified type scenarios. General request scenarios refer to users expressing a desire to play music without explicitly specifying a particular track or style (e.g., "Play music"); new song exploration scenarios refer to users explicitly requesting content they haven't heard before (e.g., "Recommend new songs"); and specified type scenarios refer to users specifying a particular genre, artist, or emotional tone (e.g., "Play some rock music").

[0137] S502, for general request scenarios, performs context-aware preference matching. The system obtains environmental parameters uploaded by the current terminal's sensors and generates context labels (such as "rainy day" or "driving"). The system searches the preference database for music style labels that have a high confidence correlation with the context label. For example, if the "driving" label has a high confidence correlation with the "fast-paced pop" style in historical data, the system will select tracks from the music library that match this style and have high historical user ratings to form a candidate list, thereby achieving environment-based adaptive recommendation.

[0138] S503, for the new song exploration scenario, performs similarity calculation based on acoustic fingerprints. To maintain stylistic consistency when recommending unfamiliar tracks, the system extracts the top N songs with the highest historical playback frequency from the user as a seed set. The system calculates the Mel-frequency cepstral coefficient (MFCC) feature vectors of all audio files in this seed set and uses their average feature vector as the user's acoustic aesthetic benchmark. Subsequently, the system searches the music library for candidate songs that the user has not played and calculates the cosine similarity between the MFCC feature vectors of the candidate songs and the user's acoustic aesthetic benchmark. The cosine similarity calculation formula is as follows:

[0139] ;

[0140] In the formula, This indicates the similarity score; Represents the user's acoustic aesthetic reference vector; Represents the acoustic feature vector of the candidate song; This represents the vector dot product operation; Represents the magnitude of a vector.

[0141] The system prioritizes adding candidate songs with similarity scores higher than a preset recommendation threshold to the recommendation list, thereby ensuring from a physical acoustic perspective that the recommended new songs match the user's listening habits. The specific algorithms for MFCC feature extraction and vectorization can be implemented using existing digital signal processing techniques by those skilled in the art, and will not be elaborated upon here.

[0142] S504 performs fine-grained filtering based on unstructured hierarchical levels for specified scenarios. When the user command specifies a main tag (such as "rock"), the system traverses the JSON hierarchical structure in the preference database. The system not only retrieves tracks containing that main tag but also checks the exclusion list of sub-genres under that main tag. If there are negative feedback records for the "heavy metal" sub-genre in the user profile (such as skipping multiple times), the system removes tracks belonging to the "heavy metal" sub-genre from the recommendation results, retaining only tracks from other sub-genres (such as "light rock" and "British rock"). This step utilizes the hierarchical nesting characteristics of JSON data to achieve fine-grained control over the recommendation granularity.

[0143] S505, the final playlist is generated and sent to the user terminal. Based on the candidate track set generated according to any of the above scenario strategies, the system sorts them according to matching score or similarity score, generating a playlist containing metadata and playback links. The system sends this playlist to the user terminal via a communication interface, where the user terminal's audio decoding and output unit loads and plays the audio stream.

[0144] See attached document Figure 7 This invention provides an apparatus for a dynamic user preference modeling music recommendation system, which is used to execute the aforementioned dynamic user preference modeling music recommendation method. The apparatus includes: a preference acquisition module 100, an unstructured preference database 200, a behavior record database 300, a matching calculation engine 400, a negative feedback control module 500, and a music execution module 600.

[0145] The preference acquisition module 100 is configured to perform bimodal data acquisition tasks. The preference acquisition module 100 receives explicit dialogue data and implicit behavioral data from the user terminal via a communication interface. The preference acquisition module 100 is configured to perform natural language processing parsing on the explicit dialogue data to extract the user's subjective intent tags; simultaneously, it extracts play, skip, and favorite operation features from the implicit behavioral data. The preference acquisition module 100 is also configured to calculate the confidence level of preference tags, assign source weights based on data source type, calculate data freshness based on timestamps, and calculate behavioral consistency based on historical recurrence frequency. Finally, it constructs an initial user preference model based on the weighted calculation results.

[0146] The unstructured preference database 200 is configured to store personalized user preference data. The unstructured preference database 200 uses a key-value pair storage structure based on JSON objects and supports dynamic data schema definition. Each user document in the unstructured preference database 200 contains hierarchical data based on genre, artist, and scene dimensions. The unstructured preference database 200 is configured to respond to write commands from the preference acquisition module 100, storing preference tags with a confidence level higher than a preset entry threshold; and to respond to update commands from the negative feedback control module 500, dynamically adding or deleting tag key-value pairs or correcting the mapping relationships between tags.

[0147] The Behavior Log Library 300 is configured to store users' historical interaction logs. It records users' actual music listening behavior data within a preset time window. Data fields include track metadata, playback duration, operation time, and corresponding contextual tags. The Behavior Log Library 300 provides data support for subsequent matching degree calculations.

[0148] The matching calculation engine 400 is configured to periodically perform consistency checks. It reads the declared preference set from the unstructured preference database 200 and the actual behavior label set from the behavior record database 300. The matching calculation engine 400 is configured to calculate label overlap and scene matching degree, and generate a comprehensive matching score of user preferences and behaviors based on a weighted summation algorithm. The matching calculation engine 400 transmits the calculated matching score to the negative feedback control module 500.

[0149] The negative feedback control module 500 is configured to execute logical decisions and trigger active interactions. The negative feedback control module 500 receives a matching score and compares it with a preset matching degree threshold. When the matching score is lower than the matching degree threshold and the time interval since the last model update exceeds a preset time limit, the negative feedback control module 500 generates an active interaction trigger signal. The negative feedback control module 500 is configured to construct a voice inquiry command containing current scene information and send it to the user terminal via a communication interface. The negative feedback control module 500 is also configured to receive real-time intent text feedback from the user terminal, perform semantic analysis on its emotional features and demand attributes, and generate update instructions for the unstructured preference database 200 based on the analysis results.

[0150] The music execution module 600 is configured to generate the final music playlist. It receives playback requests from the user terminal and parses the request type to determine if it's a general request scenario, a new song exploration scenario, or a specified type scenario. Based on the request type, the music execution module 600 calls the preference model data in the unstructured preference database 200 and the current context scenario data to execute the corresponding filtering and sorting strategies. For the new song exploration scenario, the music execution module 600 performs similarity calculations based on Mel-frequency cepstral coefficient feature vectors; for the specified type scenario, the music execution module 600 performs fine-grained filtering based on a JSON hierarchical structure. The music execution module 600 ultimately outputs a playlist containing track metadata and playback links.

[0151] It should be understood that the aforementioned preference acquisition module 100, unstructured preference database 200, behavior record library 300, matching calculation engine 400, negative feedback control module 500, and music execution module 600 refer to computer program segments that implement corresponding functions when executed by the processor.

[0152] At the hardware level, the apparatus for the dynamic user preference modeling music recommendation system described in this embodiment may include a processor, a memory, and a communication interface. The memory stores the computer program and the operational data required by the aforementioned functional modules. The processor is configured to call the computer program stored in the memory and execute the various steps of the aforementioned dynamic user preference modeling music recommendation method. Specifically, when the processor executes the computer program, it performs the following functions: controlling the preference acquisition module 100 to perform bimodal data fusion; controlling the matching calculation engine 400 to calculate the matching score; controlling the negative feedback control module 500 to initiate an active dialogue when the triggering conditions are met; and controlling the music execution module 600 to generate recommended content. The communication interface is configured to enable data communication between the apparatus and the user terminal, an external knowledge base server, and a streaming media content distribution server.

[0153] See attached document Figure 8 This invention provides a terminal for a dynamic user preference modeling music recommendation system. The terminal is configured to interact with the aforementioned music recommendation device and execute recommendation instructions. The terminal includes: an audio acquisition unit 801, an audio output unit 802, a positioning and sensor unit 803, a communication unit 804, and a processor 805.

[0154] The audio acquisition unit 801 is configured to acquire voice commands issued by the user and ambient background sounds. The audio acquisition unit 801 includes a microphone array for performing sound source localization and noise reduction processing. The audio acquisition unit 801 converts the acquired analog sound wave signals into digital audio signals and transmits the digital audio signals to the processor 805.

[0155] Audio output unit 802 is configured to play multimedia audio streams and provide speech synthesis feedback. Audio output unit 802 receives audio data from processor 805, including streaming music tracks from a server and voice query data from active interaction processes. Audio output unit 802 converts the digital audio data into sound wave signals for playback.

[0156] The positioning and sensor unit 803 is configured to monitor the terminal's geographic location, motion status, and time information in real time. The positioning and sensor unit 803 includes a Global Positioning System (GPS) module, an accelerometer, and a system clock. The GPS module acquires latitude and longitude coordinates and movement speed; the accelerometer acquires three-axis acceleration data; and the system clock provides the current timestamp.

[0157] Processor 805 is configured to execute context generation logic to generate scene labels based on data collected by positioning and sensor unit 803. Processor 805 reads the mobile speed data from the Global Positioning System module and the system clock time data. When it detects that the mobile speed is within a preset vehicle speed range and the current time is within a preset morning or evening peak period, processor 805 generates a "commuting" scene label.

[0158] The processor 805 is also configured to determine the user's movement pattern based on data from the accelerometer. When the frequency and amplitude of changes in the acceleration data are detected to match preset running or fitness rhythm characteristics, the processor 805 generates a "movement" scene label. The processor 805 associates the generated scene label with the current user ID and stores it as part of the request parameters.

[0159] Communication unit 804 is configured to establish a data connection between the terminal and the music recommendation device. Communication unit 804 encapsulates the digital audio signal processed by processor 805, scene tags, and user ID into a request data packet and sends it to the music recommendation device. Simultaneously, communication unit 804 receives music playlists, control commands, or voice interaction text returned by the music recommendation device.

[0160] The processor 805 also runs an interactive client program. This interactive client program is configured to be in standby listening mode, analyzing the audio stream input by the audio acquisition unit 801 in real time. When a preset wake-up word feature is detected, the processor 805 activates the voice command recording function.

[0161] Upon receiving an active interaction command from the music recommendation device, the processor 805 controls the audio output unit 802 to play the inquiry content. Subsequently, the processor 805 automatically activates the recording function of the audio acquisition unit 801, sets a preset recording window, acquires the user's response voice, and uploads the response voice to the music recommendation device via the communication unit 804 to complete the feedback interaction process of preference confirmation.

[0162] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for a dynamic user preference modeling music recommendation system, characterized in that, Music recommendation systems applied to both music recommendation devices and user terminals include: S10. Collect explicit dialogue data and implicit behavioral data of users through a dual-modal approach of dialogue interaction and behavior analysis, and construct an initial user preference model based on source weight, data freshness, and behavioral consistency. S20. Obtain the user's actual listening behavior data and calculate the matching degree between the user's declared preferences in the initial user preference model and the actual listening behavior data; S30. Detect whether the matching degree is lower than the matching degree threshold and meets the trigger time condition; if so, trigger active dialogue interaction and obtain the user's real-time intention feedback through voice inquiry. S40. Perform semantic analysis on the real-time intent feedback, identify emotional features or new label attributes, correct the initial user preference model based on the identification results and update it to the unstructured preference database to obtain the corrected user preference model. S50. In response to the user's playback request, generate a music playlist or recommended content based on the modified user preference model and the current context.

2. The method for a dynamic user preference modeling music recommendation system according to claim 1, characterized in that, In step S10, constructing the initial user preference model based on source weight, data freshness, and behavioral consistency includes: Preference tags are extracted from the explicit dialogue data and the implicit behavior data, and the source weights are assigned to the preference tags, wherein the source weight value of the preference tags from the explicit dialogue data is set to be greater than the source weight value of the preference tags from the implicit behavior data. The data freshness is calculated based on the timestamps generated by the preference tags, and the behavioral consistency is calculated based on the recurrence frequency of the preference tags in the historical behavior sequence. The final confidence level is calculated by considering the source weight, data freshness, and behavioral consistency. Preference labels with a final confidence level greater than the confidence level entry threshold are written into the initial user preference model.

3. The method for a dynamic user preference modeling music recommendation system according to claim 1, characterized in that, In step S20, calculating the matching degree between the user-stated preferences in the initial user preference model and the actual music listening behavior data includes: Calculate the overlap ratio between the declared preference set in the initial user preference model and the actual behavior tag set composed of the actual music listening behavior data to obtain the tag overlap degree; The scene matching degree is obtained by detecting whether the style tag of the currently played song in the actual listening behavior data is consistent with the style tag of the current context scene in the history that has a recurrence frequency exceeding a frequency threshold. The matching degree is generated based on the weighted result of the tag overlap degree and the scene matching degree.

4. The method for a dynamic user preference modeling music recommendation system according to claim 1, characterized in that, In step S30, the triggering time condition is that the time interval since the last model update exceeds the minimum update time interval; The triggering of proactive dialogue interaction includes: When the matching degree is detected to be lower than the matching degree threshold and the triggering time condition is met, a voice query content containing information about the current context scene is generated.

5. The method for a dynamic user preference modeling music recommendation system according to claim 1, characterized in that, In step S40, correcting the initial user preference model based on the recognition result includes: The emotional state slots and demand slots are extracted from the real-time intent feedback using a natural language processing model. The attributes corresponding to the demand slots are associated with the current context scenario and written into the unstructured preference database.

6. The method for a dynamic user preference modeling music recommendation system according to claim 1, characterized in that, The unstructured preference database adopts a key-value pair hierarchical structure; In step S40, identifying emotional features or new label attributes further includes: When a newly identified tag attribute is not included in the local standard tag set, the new tag attribute is marked as an unknown tag, and the description information of the unknown tag is obtained by querying an external knowledge base; The acoustic feature attribute words are extracted from the description information, the unknown tags are mapped to locally executable combined tags, and the mapping relationship between the unknown tags and the combined tags is written into the unstructured preference database.

7. The method for a dynamic user preference modeling music recommendation system according to claim 1, characterized in that, In step S50, generating a music playlist or recommended content based on the modified user preference model and the current context includes: Identify the intent type of the user's playback request; If the intent type is a general request scenario, then retrieve the music style tags associated with the current context scenario from the unstructured preference database and filter the tracks that match the music style tags; If the intent type is a new song exploration scenario, then based on the acoustic feature vectors of songs whose historical playback frequency exceeds a frequency threshold, the similarity between the acoustic feature vectors of the songs and the acoustic feature vectors of the candidate songs is calculated, and tracks with a similarity greater than the filtering threshold are filtered. If the intent type is a specified type scenario, then the specified main tag is retrieved in the hierarchical structure of the unstructured preference database, and sub-genre tracks with negative feedback records are removed.

8. The method for a dynamic user preference modeling music recommendation system according to claim 7, characterized in that, The unstructured preference database includes genre, artist, and scene fields; the process of removing tracks from sub-genres with negative feedback records includes: Check the subgenre exclusion list under the genre field and filter out tracks that fall into the subgenre exclusion list.

9. An apparatus for a dynamic user preference modeling music recommendation system, characterized in that, include: The preference collection module is configured to collect explicit dialogue data and implicit behavioral data of users through a dual-modal approach of dialogue interaction and behavior analysis, and to build an initial user preference model based on source weight, data freshness, and behavioral consistency. The matching calculation engine is configured to obtain the user's actual music listening behavior data and calculate the matching degree between the user's declared preferences in the initial user preference model and the actual music listening behavior data; The negative feedback control module is configured to detect whether the matching degree is lower than the matching degree threshold and meets the trigger time condition; if so, it triggers active dialogue interaction, obtains the user's real-time intent feedback through voice inquiry, performs semantic analysis on the real-time intent feedback, identifies emotional features or new tag attributes, and corrects the initial user preference model based on the identification results to obtain the corrected user preference model. An unstructured preference database is configured to store the initial user preference model and the revised user preference model; The music execution module is configured to generate a music playlist or recommended content in response to a user's playback request, based on the modified user preference model and the current context.

10. A terminal for a dynamic user preference modeling music recommendation system, characterized in that, include: The audio acquisition unit is configured to collect explicit dialogue data and real-time intent feedback from users. The audio output unit is configured to play music playlists or recommended content, as well as play voice-activated questions during proactive dialogue interactions. The positioning and sensor unit is configured to collect sensor data. A communication unit configured to transmit data with a music recommendation device; The processor is configured to generate the current context scene based on the sensor data; and is configured to respond to the active interaction command issued by the music recommendation device, control the audio output unit to play the voice inquiry content, and control the audio acquisition unit to acquire the real-time intent feedback within a set sampling time window.

Images