Cloud management and intelligent recommendation system for piano library

Abstract

The invention discloses a piano library cloud management and intelligent recommendation system, which belongs to the technical field of digital music resource management and recommendation and comprises a music knowledge graph construction module, a mixed recommendation module, a learning sequence generation module, a cloud database module, a copyright management module, a user interaction module and a data synchronization module. By constructing a structured knowledge graph containing difficulty classification, technical characteristics and music styles, combining a mixed recommendation algorithm of content characteristics and collaborative filtering and a progressive learning sequence generation method based on the knowledge graph, intelligent organization and personalized pushing of piano teaching resources are realized by the system, and the system can be applied to the field of piano teaching. The copyright management module guarantees legal use of resources, the data synchronization module achieves efficient synchronization of the cloud and the local, the defects in the aspects of music professional feature modeling, recommendation algorithm and learning path planning in the prior art are overcome, and the teaching adaptability and learning efficiency of recommendation are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of digital music resource management and recommendation technology, specifically to a cloud-based piano music library management and intelligent recommendation system, which is particularly suitable for intelligent organization and personalized delivery of piano teaching resources. Background Technology

[0002] Existing music recommendation systems, such as the intelligent music recommendation system disclosed in Chinese Patent CN 105718566 A (applicant: Sun Yat-sen University, application date: January 20, 2016), specifically disclose a music recommendation scheme based on a song distance network and user feedback. This system includes an initialization module, a playback module, and an adjustment module. It constructs a song distance network to calculate the correlation values ​​and distances between songs. Based on the seed songs selected by the user, it searches the network for highly correlated songs to build a personal song library, and adjusts playback weights according to the context and user feedback. The system uses user playlist data as a data source, determining the correlation values ​​between songs by calculating the co-occurrence of songs in the same playlist or different playlists of the same user, thereby constructing the song distance network. During recommendation, the system judges the playback probability based on the user's context, modifies the weights after obtaining user feedback, and periodically adjusts the song distance network and the personal song library.

[0003] However, this existing technology has the following technical problems:

[0004] First, regarding the organization of musical resources, existing technologies only construct song distance networks based on the co-occurrence relationships of user playlists, lacking in-depth modeling of the professional characteristics of piano pieces. For the specialized application scenario of piano learning, the relationships between pieces include not only stylistic similarities but, more importantly, difficulty gradients, technical characteristics, and teaching progression relationships. For example, there is a clear progression in difficulty between Czerny Etudes 599 and 849; Hanon finger exercises and Bach's Two-Part Inventions are complementary in terms of technical training objectives. However, existing technologies cannot express these specialized relationships between pieces. When a piano learner completes a piece at a certain difficulty level, the system cannot recommend appropriate pieces for the next stage based on teaching progression logic, resulting in deficiencies in the teaching suitability of the recommendations.

[0005] Secondly, in terms of recommendation algorithm design, existing technologies mainly rely on user feedback to adjust recommendation weights, which is a single collaborative filtering approach and fails to fully utilize the content features of the music itself. For piano music library management scenarios, each piece of music has rich inherent attributes, including composer, composition period, musical style, technical challenges, tempo requirements, and other multi-dimensional features. Existing technologies only consider the song's distance in the network and its contextual weight during recommendation, ignoring the discriminative role of these content features. When faced with newly added music or newly registered learners, the lack of historical interaction data leads to a cold start problem, severely degrading recommendation quality and failing to provide users with effective learning resource recommendations.

[0006] Third, regarding learning path planning, existing technologies only recommend similar pieces based on the relevance of seed songs, lacking dynamic modeling of learners' skill development and systematic planning of long-term learning paths. Piano learning is a long-term process requiring gradual progression. An ideal recommendation system should be able to automatically generate a complete learning sequence from basic to advanced levels based on the learner's current level, rather than simply recommending other pieces similar to the current piece. Existing technologies make independent decisions for each recommendation, lacking coherence and progression between recommendations, and failing to form a systematic teaching plan. When learners use the system for a long time, the recommendation results are prone to getting stuck in local optima, that is, repeatedly recommending pieces of similar difficulty without guiding learners to improve to higher levels. This deficiency limits the system's application value in professional piano teaching.

[0007] Fourth, regarding copyright management, existing technologies do not address copyright protection mechanisms for digital sheet music and instructional videos. In cloud-based music library management scenarios, a large amount of sheet music and instructional resources are stored and disseminated digitally. Ensuring the legal use of these resources and preventing unauthorized dissemination are key issues that must be addressed in practical applications. Existing technologies lack copyright tracking, usage permission control, and payment management functions for piano instructional resources, making it difficult to meet the needs of commercial operation.

[0008] Therefore, there is an urgent need for a new cloud-based piano music library management and intelligent recommendation system to solve the above-mentioned technical problems. Summary of the Invention

[0009] To address the shortcomings of existing technologies, the present invention aims to provide a cloud-based management and intelligent recommendation system for piano music libraries. By constructing a structured music knowledge graph, designing a hybrid filtering recommendation algorithm, and implementing progressive learning sequence generation, the invention solves the technical problems of insufficient modeling of professional music features, single recommendation algorithms, and lack of learning path planning in existing technologies. This improves the teaching adaptability and personalization of recommendations and introduces copyright management functions to ensure the legal use of resources.

[0010] To achieve the above objectives, the present invention provides a cloud-based piano music library management and intelligent recommendation system, which includes:

[0011] The music knowledge graph construction module is used to build a structured music knowledge graph that includes difficulty levels, technical characteristics, and musical styles.

[0012] The hybrid recommendation module combines music content features and user collaborative filtering information to generate a personalized recommendation list;

[0013] The learning sequence generation module is used to plan a progressive learning path in the music knowledge graph based on the learner's ability assessment results.

[0014] The cloud database module is used to store music metadata, digital scores, and teaching video resources;

[0015] The copyright management module is used to control access permissions for music resources and track usage records;

[0016] The user interaction module is used to receive user operation instructions, display recommendation results, and collect learning feedback data.

[0017] The data synchronization module is used to achieve incremental synchronization between cloud databases and local applications.

[0018] Furthermore, the music knowledge graph construction module includes:

[0019] The feature extraction unit is used to extract multi-dimensional features such as difficulty level, technical type, musical style, composer, and creation period from the music metadata.

[0020] The graph structure generation unit is used to represent music as nodes in a graph, and constructs directed edges based on difficulty gradient, technical similarity and style relevance to form a music knowledge graph;

[0021] Topology optimization units are used to adjust the graph structure to optimize the reachability of learning paths by utilizing connectivity analysis and path optimization methods in graph theory.

[0022] Furthermore, the hybrid recommendation module includes:

[0023] The content feature matching unit is used to calculate the matching degree between candidate music pieces and learners' ability features;

[0024] The collaborative filtering unit is used to predict the current learner's preference for candidate music pieces based on the historical learning records of similar learners;

[0025] The hybrid fusion unit is used to fuse content matching scores and collaborative filtering scores through adaptive weights to generate a comprehensive recommendation score.

[0026] Furthermore, the learning sequence generation module includes:

[0027] The competency assessment unit is used to quantitatively assess a learner's current competency level based on their historical learning records and test results.

[0028] The path search unit is used to search for learning paths that satisfy the increasing difficulty constraint in the music knowledge graph, starting from the music node corresponding to the learner's current ability.

[0029] The sequence optimization unit is used to optimize the generated learning path, balancing path length, difficulty gradient, and content diversity.

[0030] Furthermore, the cloud database module includes:

[0031] Metadata storage unit, used to store structured metadata of the music, including attributes such as difficulty, style, and technical characteristics;

[0032] The file storage unit is used to store image files of digital sheet music and multimedia files of teaching videos;

[0033] The index building unit is used to create a multidimensional index for music metadata, supporting fast retrieval.

[0034] Furthermore, the copyright management module includes:

[0035] The access control unit is used to set access permissions for different music resources based on the user's subscription status and payment history;

[0036] A tracking unit is used to record the time, frequency, and duration of a user's access to music resources;

[0037] The watermark embedding unit is used to embed invisible digital watermarks in digital sheet music and instructional videos to identify the user.

[0038] Furthermore, the user interaction module includes:

[0039] The query interface unit is used to receive users' music search requests and filter criteria;

[0040] The display rendering unit is used to present the recommendation results to users in a list or knowledge graph visualization format;

[0041] The feedback collection unit is used to collect feedback data such as user ratings of recommended music, learning completion rate, and learning time.

[0042] Furthermore, the data synchronization module includes:

[0043] The difference detection unit is used to compare the data versions between the cloud database and the local cache to identify the data items that need to be synchronized.

[0044] Incremental transmission units are used to transmit only the data content that has changed, thereby reducing network bandwidth consumption.

[0045] The conflict resolution unit is used to handle conflicts when the same data item is modified simultaneously in the cloud and locally, and merges them using a strategy that prioritizes timestamps or requires user confirmation.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] First, by constructing a structured music knowledge graph, this invention achieves deep modeling of the professional characteristics of piano pieces. The knowledge graph structurally expresses multi-dimensional attributes such as difficulty levels, technical features, and musical styles of the pieces, and uses directed edges to represent the learning progression relationships between pieces. Compared to existing technologies that rely solely on song distance networks based on co-occurrence relationships, the knowledge graph of this invention more accurately reflects the professional knowledge in piano teaching, making the recommendation results more consistent with teaching logic. Experimental data shows that the knowledge graph-based recommendation system improves the teaching suitability score by 18% to 25% compared to traditional methods.

[0048] Secondly, by designing a hybrid filtering recommendation algorithm, this invention effectively solves the cold start problem and improves recommendation accuracy. The hybrid recommendation module utilizes both the content features of the music and user collaboration information. When faced with new music or new users, it can rely on content features for recommendations, avoiding the performance degradation of pure collaborative filtering methods in data-sparse situations. The adaptive weight fusion mechanism dynamically adjusts the weights of content matching and collaborative filtering based on data availability, maintaining high recommendation quality in different scenarios. Comparative experiments show that the recommendation accuracy of the hybrid recommendation algorithm is 12% to 20% higher than that of the single collaborative filtering method, with a more significant improvement in cold start scenarios.

[0049] Third, by generating progressive learning sequences, this invention provides learners with a systematic learning path planning approach. The learning sequence generation module searches the music knowledge graph for paths that satisfy the increasing difficulty constraint and balances path length, difficulty gradient, and content diversity through an optimization algorithm. The generated learning sequences guide learners to gradually improve their abilities while maintaining the richness of the learning content. Compared to existing technologies that recommend similar tracks independently each time, the sequence generation method of this invention has stronger systematicity and long-term planning capabilities, improving user learning efficiency by approximately 15% to 22%.

[0050] Fourth, by introducing a copyright management module, this invention ensures the legal use of digital sheet music and instructional video resources. The copyright management module implements functions such as access control, usage tracking, and digital watermark embedding, which not only prevents the unauthorized dissemination of resources but also provides copyright holders with usage data analysis support, meeting the needs of commercial operation. This function fills the gap in copyright protection in existing music recommendation systems and provides a complete technical solution for the digital management of piano teaching resources.

[0051] Fifth, through an efficient synchronization mechanism between the cloud database and local applications, this invention enables convenient access to massive amounts of teaching resources. The data synchronization module adopts an incremental transmission strategy, synchronizing only changed data content, significantly reducing network bandwidth consumption and improving user experience. Even under poor network conditions, users can smoothly use locally cached resources, and data synchronization is automatically completed after the network is restored, ensuring the availability and stability of the system. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the overall architecture of the piano music library cloud management and intelligent recommendation system of the present invention;

[0053] Figure 2 This is a schematic diagram of the structure of the music knowledge graph construction module of the present invention;

[0054] Figure 3 This is a schematic diagram of the hybrid recommendation module of the present invention;

[0055] Figure 4 This is a schematic diagram of the learning sequence generation module of the present invention;

[0056] Figure 5 This is an example structural diagram of the musical knowledge graph of the present invention. Detailed Implementation

[0057] Please refer to the attached document. Figure 1-5 The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0058] like Figure 1As shown, the piano music library cloud management and intelligent recommendation system provided by this invention includes seven core modules: a music knowledge graph construction module, a hybrid recommendation module, a learning sequence generation module, a cloud database module, a copyright management module, a user interaction module, and a data synchronization module. These modules work together to achieve intelligent organization and personalized delivery of piano teaching resources.

[0059] The overall workflow of the system is as follows: First, the music knowledge graph construction module obtains music metadata from the cloud database module, extracts multi-dimensional features such as difficulty, style, and technical characteristics, and constructs a structured knowledge graph. This graph uses music pieces as nodes and learning progression relationships as edges, forming a graph structure that expresses piano teaching knowledge. Then, when a user initiates a recommendation request through the user interaction module, the hybrid recommendation module combines the content features of the music pieces with collaborative filtering information from other users to calculate the recommendation score of candidate music pieces and generate a preliminary recommendation list. Next, the learning sequence generation module searches the music knowledge graph for learning paths that satisfy the increasing difficulty constraint based on the user's ability assessment results, organizing the recommendation list into a systematic learning sequence. Throughout the process, the copyright management module controls the user's access permissions to different music resources, records usage behavior, and embeds digital watermarks. Finally, the user interaction module displays the recommendation results to the user and collects user learning feedback data, which is used for subsequent recommendation optimization. The data synchronization module works continuously in the background to ensure data consistency between the cloud database and the user's local application, employing an incremental transmission strategy to reduce network overhead.

[0060] The entire system adopts a modular design, with each module communicating through standardized data interfaces. The music knowledge graph serves as the core data structure, shared and accessed by multiple modules. The hybrid recommendation module utilizes node features in the graph for content matching, while the learning sequence generation module uses the edge structure of the graph for path searching. This design ensures both high system cohesion and low coupling between modules, facilitating system maintenance and functional expansion.

[0061] like Figure 2 As shown, the music knowledge graph construction module is one of the core innovations of this invention. This module includes a feature extraction unit, a graph structure generation unit, and a topology optimization unit.

[0062] The feature extraction unit extracts multi-dimensional features from the music metadata stored in the cloud database. Preferably, for each piece of music in the music library, the feature extraction unit extracts the following attributes: difficulty level, technical type, musical style, composer, composition period, rhythmic features, harmonic features, performance duration, and recommended learning stage.

[0063] Specifically, the difficulty level adopts a ten-level grading system, from Level 1 (beginner) to Level 10 (professional performance level). This grading is based on internationally authoritative piano examination systems and combined with the annotations of music education experts. Technical type tags include, but are not limited to: fingering training, scale exercises, arpeggio exercises, chord training, octave training, vibrato training, staccato training, legato training, pedal application, polyphonic techniques, and fast runs. Each piece can be associated with multiple technical type tags. Musical style classifications include Baroque, Classicism, Romanticism, Impressionism, Modernism, Jazz, and Pop, reflecting both historical periods and contemporary popular music styles. Composer information records the composer of the work, and the composition period indicates the year or approximate period of composition. Rhythmic features extract the tempo markings (e.g., Allegro, Andante) and time signature information, while harmonic features indicate the main harmonic techniques used (e.g., major / minor scales, seventh chords, altered chords). Performance duration records the standard performance duration range, and recommended learning stages indicate which stage of learning the piece is suitable for practice.

[0064] The feature extraction unit organizes these multidimensional features into structured feature vectors. Preferably, for each piece of music... Its eigenvectors are represented as:

[0065] ,

[0066] in: For music The complete feature vector; The difficulty feature sub-vector is represented by one-hot encoding, with a dimension of 10; For technical feature sub-vectors, multi-label encoding is used to represent technical types. The dimension is the total number of technical labels, and the value is 0 or 1. 1 indicates that the music contains the corresponding technical type. The style feature sub-vectors are represented by one-hot encoding to indicate the main music styles, with the dimension being the total number of style categories. The composer features are represented by embedding vectors with a dimension of 64, obtained through a pre-trained composer embedding model; To represent the period characteristics, normalized values ​​of the creation year are expressed as scalars or small-dimensional vectors.

[0067] In the implementation process, the feature extraction unit first reads the metadata records of the music from the database, and then encodes and transforms each attribute according to a predefined feature template. For discrete attributes such as difficulty level and musical style, one-hot encoding is used to generate sparse vectors; for multi-label attributes such as technical type, multi-label binary vectors are generated; for high-cardinality discrete attributes such as composer, pre-trained embedding vectors are used for dimensionality reduction representation. The embedding vectors are trained on a large-scale music dataset and can capture the similarity of compositional styles among different composers. Finally, the feature extraction unit concatenates the sub-vectors into a unified feature vector and performs normalization to ensure that features of different dimensions are comparable on a numerical scale.

[0068] The graph structure generation unit transforms the extracted musical feature vectors into a graph structure representation of a knowledge graph. In this graph structure, each musical piece corresponds to a node in the graph, and the node's attributes are the feature vector of that musical piece. The learning progression relationships, technical similarity relationships, and style association relationships between musical pieces are expressed through directed or undirected edges.

[0069] Specifically, the music knowledge graph constructed by the graph structure generation unit. ,in For a set of nodes, each node Represents a piece of music; Let be a set of edges, and the edges are divided into three types: difficulty progression edges, technical similarity edges, and style-related edges.

[0070] Construction of difficulty progression edges: Difficulty progression edges represent the sequential learning relationship between two pieces of music. Preferably, when the music... The difficulty level is higher than that of the music. When the two pieces are one or two levels higher and overlap in technical type, from the node... To the node Construct a directed edge This indicates that the learner has mastered the piece of music. After that, you can try learning the music. The weights of difficulty-advancing edges are calculated based on the difficulty gradient and technical accessibility, using the following formula:

[0071] ,

[0072] in: The weight of the edge representing the increasing difficulty level, with a value ranging from 0 to 1; For music The difficulty level is an integer ranging from 1 to 10; For music The difficulty level is an integer ranging from 1 to 10; The Gaussian kernel width parameter for the difficulty gradient is used to control the tolerance range of difficulty differences. A preferred value is 0.5 to 1.0; in this embodiment, it is 0.7. For music and The intersection of the technical type tags indicates the technical types that are commonly contained in the two pieces of music; The union of technology type labels; This indicates the number of elements in the set.

[0073] The first term of the above formula This is a Gaussian kernel function used to quantify the reasonableness of the difficulty gradient. When When the difficulty level increases by one level, this item is set to 1, indicating the most reasonable difficulty gradient. When the difficulty difference deviates from one level, the value of this item decreases as the deviation increases, reflecting the principle of gradual progression in teaching. (Second item) is the Jaccard similarity coefficient, used to measure the overlap in technical types between two pieces of music; a larger coefficient indicates better technical compatibility. Multiplying the two terms yields an edge weight that comprehensively considers the difficulty gradient and technical compatibility. This weight is used as the basis for path optimization in subsequent learning path search.

[0074] To more clearly illustrate the construction process of difficulty progression edges, a specific example is provided below. Assume the music library contains two pieces of music... This is Czerny Etude 599, No. 1, with a difficulty level of 3 and technical type tags of {fingering exercise, scale exercise, legato exercise}; the piece... This is Czerny Etude 849, No. 1, with a difficulty level of 5 and technical type tags of {fingering exercises, scale exercises, fast runs}. Based on the above formula, the difficulty gradient term is first calculated, since... Substituting into the Gaussian kernel function, we get Then, the technical continuity term is calculated. The intersection of the technical type labels of the two pieces is {fingering training, scale practice}, with 2 elements; the union is {fingering training, scale practice, legato training, fast runs}, with 4 elements. Therefore, the Jaccard coefficient is... Ultimately, the weights of the difficulty-advancing edges are: This weight represents the weight from the musical piece. Learned music The transition is somewhat reasonable, but because the difficulty jumps across two levels, its weight is relatively low. Through this mechanism, the system can automatically identify learning leaps with excessively large difficulty gradients and assign them lower priority during recommendation.

[0075] Construction of technically similar edges: Technically similar edges represent the similarity between two pieces of music in terms of technical training objectives, and are used to recommend practice pieces with complementary or reinforcing relationships. Preferably, when the difficulty levels of two pieces of music are the same or differ by no more than one level, and the overlap of their technical type labels exceeds a preset threshold, an undirected edge is constructed between the corresponding nodes. The weight calculation formula for technically similar edges is as follows:

[0076] ,

[0077] in: The weight of the edge representing technological similarity, with a value ranging from 0 to 1; The cosine similarity of the technical feature vectors is calculated using the following formula: ,in Represents the vector dot product. Representing vectors Norm; second term This is a penalty for differences in difficulty. When the two pieces of music are at the same difficulty level, this penalty is 1. As the difference in difficulty increases, this penalty decreases linearly. When the difference in difficulty reaches 10 levels, this penalty is 0.

[0078] The purpose of technical similarity edges is to help learners expand their technical reserves horizontally at the same level of difficulty. For example, two etudes containing octave training techniques, even if composed by different composers and with different styles, can be linked through technical similarity edges, allowing learners to reinforce the same technical points through diverse works. The weight of this edge considers both the similarity of technical types and the matching of difficulty levels, ensuring that the recommended pieces, while maintaining consistent technical goals, do not reduce learning effectiveness due to a mismatch in difficulty.

[0079] Construction of style-related edges: Style-related edges represent the similarity of two musical pieces in terms of musical style, and are used to recommend tracks with consistent or diverse styles to learners. Preferably, when two musical pieces belong to the same musical style category, an undirected edge is constructed between the corresponding nodes, with a fixed weight of 0.8. Furthermore, for musical pieces with different styles but similar harmonic or rhythmic features, style-related edges with lower weights can also be constructed, with the weight calculated based on the similarity of harmonic and rhythmic features.

[0080] The introduction of style-related edges enables the knowledge graph to support recommendations based not only on teaching logic but also on musical aesthetics. Learners may wish to delve deeper into works of a specific style or explore diverse musical styles; the system's style-related edge structure flexibly caters to both needs.

[0081] Graph structure fusion: The graph structure generation unit integrates the above three types of edges into a unified knowledge graph. Preferably, the final edge weight matrix... The weighted fusion yielded:

[0082] ,

[0083] in: The merged edge weight matrix has dimensions of . , This represents the total number of songs in the music library. The weight matrix for edges of increasing difficulty. This is the weight matrix of the edges with similar technologies. This is the weight matrix for style-related edges; , , To integrate the weighting coefficients, satisfy the following conditions: Preferably , , This configuration assigns the highest weight to the difficulty progression relationship, reflecting the primary principle of teaching as gradual progression, while also taking into account technical similarity and stylistic consistency.

[0084] Through the aforementioned fusion mechanism, the graph structure generation unit constructs a knowledge graph that comprehensively expresses the relationships between various musical pieces. This graph contains both directed learning path information and undirected similarity information, providing rich structured knowledge support for subsequent recommendation and sequence generation.

[0085] The topology optimization unit uses connectivity analysis and path optimization methods from graph theory to optimize the constructed knowledge graph, thereby improving the accessibility of learning paths and the quality of recommendations.

[0086] Specifically, the topology optimization unit performs the following optimization operations:

[0087] First, connected component analysis. The topology optimization unit performs connected component analysis on the knowledge graph to identify isolated nodes or weakly connected subgraphs. Isolated nodes indicate that some pieces of music lack connections with other pieces in the graph. This may be due to the piece's features being too unique or the graph construction rules being too strict. For isolated nodes, the topology optimization unit adopts a strategy of relaxing constraints, such as increasing the tolerance range of the difficulty gradient or lowering the threshold of technical similarity, attempting to establish at least one connecting edge for the isolated node to ensure that all pieces of music can be included in the recommendation scope.

[0088] Second, path redundancy detection. The topology optimization unit detects whether there are too many redundant paths in the graph, i.e., a large number of different paths from the same starting point to the same ending point. Too many redundant paths increase the computational complexity of generating learning sequences and may lead to overly scattered recommendation results. For detected redundant paths, the topology optimization unit retains the highest-weighted paths and deletes the lower-weighted redundant paths based on the path length and the average weight of the edges on the path, thereby simplifying the graph structure.

[0089] Third, shortest path pre-computation. To accelerate path search in the subsequent learning sequence generation module, the topology optimization unit pre-computs the shortest paths between all node pairs in the knowledge graph. Preferably, the Floyd-Warshall algorithm or the Johnson algorithm is used to calculate the shortest paths across all sources, and the results are stored as a shortest path matrix. ,in Indicates from node To the node The shortest path length is defined as the minimum sum of the reciprocals of the weights of the edges on the path; edges with larger weights have shorter distances, which aligns with the goal of learning path optimization. Pre-computed shortest path information is cached and queried directly at runtime, significantly improving system response speed.

[0090] Fourth, community detection and stratification. The topology optimization unit performs community detection on the knowledge graph, identifying subsets of music pieces with dense internal connections. These subsets typically correspond to specific technical topics or stylistic genres. Based on the community structure, the system can provide learners with stratified recommendations; for example, it can first recommend music pieces within the same community to deepen topical learning, and then recommend music pieces across communities to broaden learning scope. Community detection employs the Louvain or Leiden algorithm, which can efficiently partition communities on large-scale graphs and output modularity metrics to evaluate the quality of the community structure.

[0091] Through the processing of topology optimization units, the structural quality of the music knowledge graph is significantly improved, providing a high-quality knowledge infrastructure for the hybrid recommendation module and the learning sequence generation module.

[0092] like Figure 3 As shown, the hybrid recommendation module combines the content features of the music with user collaborative filtering information to generate a personalized recommendation list. This module includes a content feature matching unit, a collaborative filtering unit, and a hybrid fusion unit.

[0093] The content feature matching unit calculates the matching degree between candidate musical pieces and learner ability features. Preferably, the learner's ability features are represented as ability vectors. The vector contains the learner's current difficulty level, the types of skills mastered, and musical style preferences.

[0094] The learner's ability vector is constructed as follows: the difficulty level feature uses a ten-level coding system, the same as the difficulty level of the music, and is determined based on the average difficulty of the music the learner has recently learned; the skill mastery feature uses a multi-dimensional vector, with each dimension corresponding to a skill type, and a value from 0 to 1, representing the degree of mastery of that skill, calculated by statistically analyzing the frequency of each skill type in the learner's historical learning records and the completion rate of the corresponding music; the style preference feature also uses a multi-dimensional vector, with each dimension corresponding to a musical style, and the value representing the learner's preference for that style, calculated by combining the learner's explicit ratings (such as ratings for recommended music) and implicit feedback (such as learning time and number of repeated practice sessions).

[0095] Content feature matching unit calculates candidate musical pieces With learners Matching score The calculation formula is:

[0096] ,

[0097] in: Candidate music The content matching score, ranging from 0 to 1; , , The weight coefficients for the three dimensions satisfy the following conditions: Preferably , , ; The difficulty matching function is defined as follows: ,in For learners The current level of difficulty, The desired increase in difficulty is preferably between 0 and 1. To match the tolerance parameter, a value between 0.8 and 1.2 is preferred; in this embodiment, 1.0 is used. For learners The degree vector of technical mastery For learners Style preference vector; This is the cosine similarity function.

[0098] Difficulty matching function The design reflects the zone of proximal development theory, which recommends music pieces slightly above a learner's current level to promote skill development without being too difficult and causing frustration. When the difficulty level is within the learner's optimal challenge range, the matching function reaches its maximum value. As the difficulty deviates from this range, the matching score decreases with increasing deviation. This function allows content recommendations to intelligently guide learners towards higher levels of difficulty, rather than keeping them in their comfort zone.

[0099] To more clearly illustrate the calculation process of content feature matching, a specific example is provided below. Let's assume the learner... The current difficulty level is level 5. The technical mastery vector has a value of 0.7 in the chord training dimension, a value of 0.4 in the pedal usage dimension, and a value of 0.5 in all other dimensions. The style preference vector has a value of 0.9 in the Romantic dimension, a value of 0.6 in the Classical dimension, and a value of 0.3 in all other dimensions. Current candidate pieces. The difficulty level is 6, the technical types include chord training and pedal usage, and the musical style is Romantic. First, calculate the difficulty matching criteria, setting... , Substituting into the formula, we get , indicating the music The difficulty level perfectly matches the learners' expected improvement level. Then, the technical similarity term is calculated, and the musical piece... The technical feature vector has a value of 1 in the chord training and pedal usage dimensions, and 0 in other dimensions. The cosine similarity is calculated with the learner's technical mastery vector, yielding a value of 0.78. Finally, the style similarity term is calculated for the musical piece. The style vector has a value of 1 in the Romanticism dimension and 0 in other dimensions. The cosine similarity between this vector and the learner's style preference vector is calculated to be 0.85. The three terms are then weighted and summed. This score represents the musical piece. It is highly relevant to learners in terms of content and has great recommendation value.

[0100] Collaborative filtering units predict a current learner's preference for candidate music pieces based on the historical learning records of similar learners. The core idea of ​​collaborative filtering is that similar users have similar preferences. By finding other learners with similar learning behaviors to the current learner, and using their ratings or learning completion rates of candidate music pieces, the current learner's interest in that music piece can be predicted.

[0101] Specifically, the collaborative filtering unit first constructs a user-music interaction matrix. The rows of the matrix represent learners, the columns represent musical pieces, and the matrix elements... Indicate learner music The score or learning completion rate is represented. In practical applications, this matrix is ​​highly sparse because each learner only learns a small portion of the music in the library.

[0102] The collaborative filtering unit uses matrix factorization to perform a low-rank decomposition of the interaction matrix, approximating it as the product of two low-dimensional matrices:

[0103] ,

[0104] in: The user-music interaction matrix has the following dimensions: , The total number of learners The total number of musical pieces; The user latent factor matrix has dimensions of . Each line For learners The latent factor vector; The latent factor matrix of the musical piece has a dimension of . Each line For music The latent factor vector; The hidden factor dimension is preferably set to 20 to 50, and in this embodiment, it is set to 30.

[0105] Matrix factorization is trained by minimizing the reconstruction error, and the objective function is:

[0106] ,

[0107] in: The set of indices for the observed user-music interactions, i.e., the matrix. The positions of non-zero elements; For learners The latent factor vector and the music The dot product of the latent factor vectors is used to predict the score; This is the regularization coefficient, used to prevent overfitting. A preferred value is between 0.01 and 0.1; in this embodiment, it is 0.05. It is the Frobenius norm, defined as the square root of the sum of the squares of all elements of the matrix.

[0108] The above optimization problem is solved using either Alternating Least Squares (ALS) or Stochastic Gradient Descent (SGD). After training, for the learner... and candidate music The collaborative filtering prediction score is calculated as follows:

[0109] ,

[0110] This score represents the current learner's preference for the piece of music based on the preferences of similar learners. The expected level of interest.

[0111] To enhance the effectiveness of collaborative filtering, the collaborative filtering unit incorporates a time decay mechanism and confidence weighting. The time decay mechanism applies exponentially decaying weights to historical interaction data, assigning higher weights to recent learning behaviors to reflect the dynamic changes in learners' interests. Confidence weighting adjusts weights based on the reliability of the interaction data. For example, for music pieces with high learning completion rates, their corresponding interaction data is considered more reliable and given higher weights; while for music pieces that have only been briefly browsed without deep learning, their interaction data receives lower weights.

[0112] The hybrid fusion unit combines content matching scores and collaborative filtering scores using adaptive weights to generate a final comprehensive recommendation score. The fusion formula is:

[0113] ,

[0114] in: Candidate music The final recommended score; The fusion weight for content matching score; The fusion weights for collaborative filtering scores; .

[0115] The innovation of the hybrid fusion unit lies in the adaptive adjustment of the fusion weights. Preferably, the fusion weights are dynamically adjusted based on data availability, and the calculation formula is as follows:

[0116] ,

[0117] ,

[0118] in: For learners The amount of historical interaction data, i.e. the number of musical pieces that the learner has learned; This is a smoothing parameter used to adjust the speed of weight switching. It is preferably set to a value between 10 and 30, and in this embodiment, it is set to 20.

[0119] The design logic of this adaptive mechanism is as follows: when the learner is a new user, there is very little historical interaction data ( When it is close to 0, Approaching 0 When the value approaches 1, recommendations primarily rely on content feature matching, effectively addressing the cold start problem; as learners use the system and accumulate interaction data, Increase As the weight of collaborative filtering gradually increases, personalized collaborative information is gradually incorporated into recommendations; when... Much larger hour, Approaching a value of 1, the recommendation primarily relies on collaborative filtering, fully leveraging collective intelligence to provide accurate recommendations. This dynamic weight adjustment mechanism enables the hybrid recommendation module to maintain high recommendation quality across different data scenarios, addressing both the challenges of a cold start and the advantages of collaborative filtering when data is plentiful.

[0120] The hybrid fusion unit calculates the final recommendation score for all candidate pieces, sorts them from highest to lowest score, and selects the top-ranked pieces. The first piece of music is output as a recommended list, in which The recommended list length can be set by the user or by the system default, with a preferred value of 10 to 50.

[0121] like Figure 4 As shown, the learning sequence generation module further organizes the recommendation list generated by the hybrid recommendation module into a systematic, progressive learning sequence. This module includes a capability assessment unit, a path search unit, and a sequence optimization unit.

[0122] The competency assessment unit quantitatively evaluates learners' current competency level based on their historical learning records and test results. The competency assessment includes not only determining the level of difficulty but also evaluating the mastery of each technical dimension.

[0123] Specifically, the ability assessment unit calculates learners Comprehensive ability score The calculation formula is:

[0124] ,

[0125] in: For learners The overall ability score is aligned with the difficulty level scale. For learners A collection of musical pieces that have already been learned; The number of elements in the set; For music Difficulty level; For learners of music The learning completion rate, ranging from 0 to 1, is calculated based on learning time, number of practice sessions, and self-evaluation.

[0126] This formula quantifies a learner's current ability by weighting the difficulty of the pieces they have already learned. Pieces with higher completion rates have a greater weight in the average, reflecting the impact of actual mastery on ability assessment. The ability assessment unit also calculates a mastery vector for each technical dimension; each component of this vector is obtained by statistically analyzing the learner's average completion rate on pieces related to that technical type.

[0127] The results of the ability assessment are used to determine the starting point for generating the learning sequence. Preferably, the starting point for the learning sequence is selected from the difficulty level in the knowledge graph. The nearest nodes, these nodes form the starting set. Sequence generation will start from the starting set and search for paths to higher levels of difficulty.

[0128] The path search unit searches the music knowledge graph for learning paths that satisfy the increasing difficulty constraint. The goal of path search is to find a path starting from the set of origins. Starting from there, passing through several intermediate nodes, we arrive at the final set corresponding to the target difficulty level. The path is such that there are progressive difficulty edges or technically similar edges between adjacent nodes on the path.

[0129] Specifically, the path search unit employs an improved Dijkstra's algorithm or an A* search algorithm for path search. The core of the search algorithm is defining a cost function for the path, which comprehensively considers path length, difficulty gradient, and edge weights. Preferably, the path cost is defined as:

[0130] ,

[0131] in: For path The cost; A path is represented as a series of directed edges. sequence; For the edge The weights are used to determine the learning transition. Higher weights indicate a more reasonable learning transition. Therefore, the reciprocal of the weight is used in the path cost, with higher weights corresponding to lower costs. This represents the length of the path, i.e., the number of edges in the path. This is a penalty coefficient for path length, used to balance path quality and path length. It is preferably set to 0.1 to 0.5, and in this embodiment, it is set to 0.2.

[0132] The pathfinding algorithm starts from each node in the starting set and expands the path according to the principle of minimum cost until it reaches the ending set or a preset maximum path length. During the search, the algorithm maintains a priority queue, where the elements are the currently explored partial paths, sorted by path cost. Each time, the path with the minimum cost is selected from the queue for expansion. During expansion, the algorithm attempts to add all outgoing edges from the terminal node of the current path to the existing path, generating new candidate paths, which are then added to the priority queue. To avoid path loops, the algorithm records the nodes visited on each path and skips visited nodes during expansion.

[0133] The path search unit outputs several paths with the lowest cost, which serve as candidate learning sequences. Preferably, the top 5 to 10 paths are output to provide diverse sequence selection.

[0134] The sequence optimization unit optimizes the candidate learning sequences output by the path search unit, balancing path length, difficulty gradient, and content diversity, and finally selects the optimal learning sequence to recommend to the user.

[0135] The sequence optimization unit defines a comprehensive scoring function for the sequence; a higher comprehensive score indicates better sequence quality. The scoring function is:

[0136] ,

[0137] in: For learning sequence The overall quality score; A difficulty smoothness score is assigned to measure the reasonableness of the difficulty gradient in the sequence; Scoring for content diversity measures the richness of musical pieces in the sequence in terms of style and technical type; Rate interest to measure the match between the music in the sequence and the learner's interests; , , Let be the weighting coefficient, satisfying Preferably , , .

[0138] Difficulty Smoothness Rating The calculation formula is:

[0139] ,

[0140] The formula calculates the deviation of the difficulty difference between adjacent nodes in the path from the ideal gradient (a one-level increase in difficulty). A smaller deviation indicates a smoother difficulty transition and a higher score. When the difficulty of all adjacent nodes increases by exactly one level... To achieve optimal results.

[0141] Content diversity rating This measures the richness of musical styles and techniques within a sequence. Preferably, the ratio of the set's cardinality to the path length is used as a measure of diversity.

[0142] ,

[0143] in: The union of the musical styles involved in all the songs in the path; The total number of musical styles in the music library; The union of the technical types involved in all the music in the path; This represents the total number of technology types. This score measures the breadth of styles and technologies covered by the sequence; the broader the coverage, the higher the score.

[0144] Interest rating The degree of matching between the music in the sequence and the learner's interests and preferences is measured and calculated as the average content matching score of all music in the sequence:

[0145] ,

[0146] in, For music A content matching score is calculated in the hybrid recommendation module. This score ensures that the generated learning sequences not only conform to the teaching logic but also align with the learner's personal interests, thereby enhancing learning motivation.

[0147] The sequence optimization unit calculates the overall quality score of all candidate sequences and selects the sequence with the highest score as the final recommended learning sequence for the user. Preferably, the system simultaneously displays the top three sequences with the highest quality scores, allowing the user to choose according to their needs, providing a degree of flexibility.

[0148] Through the collaborative work of the ability assessment unit, path search unit, and sequence optimization unit, the learning sequence generation module realizes personalized progressive learning path planning, providing learners with a systematic learning plan from the current level to a higher level. Compared with the method of recommending similar tracks independently each time, it has significant teaching advantages.

[0149] The cloud-based database module is responsible for storing and managing all data resources in the piano repertoire library, including music metadata, digital scores, and instructional videos. This module includes metadata storage units, file storage units, and index building units.

[0150] The metadata storage unit uses a relational database or document database to store the structured metadata of the musical pieces. The metadata for each piece includes fields such as piece identifier, name, composer, difficulty level, technical type tag, musical style, composition period, sheet music file path, instructional video file path, and copyright information. The metadata storage unit supports CRUD operations on metadata and provides transaction management to ensure data consistency.

[0151] The file storage unit uses object storage services to store digital sheet music image files and instructional video multimedia files. Digital sheet music is stored in PDF or image format, while instructional videos are stored in MP4 or other compressed formats. The file storage unit supports chunked uploads of large files and resume uploads from where interrupted, improving upload efficiency and reliability. Simultaneously, the file storage unit implements version control, retaining historical versions for future reference when sheet music or videos are updated.

[0152] The index building unit creates multi-dimensional indexes for music metadata, supporting fast retrieval. Preferably, the indexes include difficulty level indexes, composer indexes, musical style indexes, and technical type indexes. Through these indexes, users can quickly find music that meets their requirements based on different filtering criteria. The index building unit uses efficient data structures such as B-trees or hash tables to implement the indexes, achieving a query time complexity of logarithmic or constant time, enabling fast retrieval of large-scale music libraries.

[0153] The cloud database module interacts with other modules through standardized data interfaces. The music knowledge graph construction module reads metadata from the metadata storage unit to construct the graph. The hybrid recommendation module queries metadata and indexes to obtain candidate music information. The copyright management module accesses the copyright information field for access control. The data synchronization module reads the database change log to identify the data items that need to be synchronized.

[0154] The copyright management module is responsible for controlling access permissions to musical resources and tracking usage records to ensure the legal use of digital sheet music and instructional videos. This module includes an access control unit, a usage tracking unit, and a watermark embedding unit.

[0155] The access control unit sets access permissions for different music resources based on users' subscription status and payment history. Music resources in the system are divided into free and paid resources. Free resources are open to all users, while paid resources are only available to subscribers or users who make a one-time purchase of the resource. The access control unit maintains a user permission table, recording each user's access permissions for each paid song. When a user requests access to a resource, the access control unit queries the permission table to determine if the user has permission. If not, the request is rejected, and the user is prompted to purchase or subscribe.

[0156] The tracking unit records users' access times, frequency, and duration of access to musical resources. This data is used for copyright statistics, recommendation algorithm optimization, and user behavior analysis. Preferably, the tracking unit generates a usage record each time a user accesses sheet music or watches an instructional video. The record includes fields such as user identifier, music identifier, access timestamp, access type (browsing sheet music or watching video), and usage duration. Usage records are stored in a log database, supporting subsequent aggregation analysis and report generation. Copyright holders can use the data provided by the tracking unit to understand the dissemination and popularity of their works, providing a basis for copyright pricing and licensing decisions.

[0157] The watermark embedding unit embeds invisible digital watermarks into digital sheet music and instructional videos to identify the user. Digital watermarking technology embeds specific information into the pixels or frequency domain of an image or video. This information is invisible to the human eye or visual perception but can be detected and extracted by specialized algorithms. Preferably, the watermark embedding unit generates watermark information containing the user's identifier and access time in real time when the user downloads or views the sheet music online, and embeds it into the sheet music image. For instructional videos, the watermark embedding unit embeds watermark frames in the video stream or modifies the video's frequency domain features. The embedded watermark is robust and can resist a certain degree of image processing operations (such as cropping, scaling, and compression) and video editing operations, ensuring that the source user can still be traced through the watermark even if the resource is illegally disseminated. The watermark detection function supports the extraction of watermark information from suspicious sheet music images or videos to identify the source of unauthorized dissemination.

[0158] The copyright management module provides a comprehensive copyright protection mechanism for the digital management of piano teaching resources through a combination of access control, usage tracking, and watermark embedding. This not only safeguards the legitimate rights and interests of copyright holders but also provides users with a convenient resource access experience, promoting the healthy development of the piano teaching resource ecosystem.

[0159] The user interaction module is responsible for receiving user operation commands, displaying recommendation results, and collecting learning feedback data. This module includes a query interface unit, a display rendering unit, and a feedback collection unit.

[0160] The query interface unit provides a user interface that receives music search requests and filtering criteria from users. Users can initiate queries by entering keywords (such as music title or composer name), selecting filtering criteria (such as difficulty level range, musical style, or technical type), or directly requesting system recommendations. The query interface unit parses user input, generates structured query requests, and calls the corresponding backend modules for processing. For keyword searches, the query interface unit calls the index retrieval function of the cloud database module; for recommendation requests, the query interface unit calls the hybrid recommendation module and the learning sequence generation module.

[0161] The display rendering unit presents the recommendation results to users in an intuitive way. Preferably, the display rendering unit supports two display modes: list mode and knowledge graph visualization mode. List mode displays the recommended music in a list format, with each list item containing information such as the music's name, composer, difficulty level, and recommendation score. Users can click on list items to view music details, play previews, or access sheet music and instructional videos. Knowledge graph visualization mode displays the music knowledge graph graphically, with nodes representing music pieces and edges representing the relationships between them. Users can interactively browse the graph, click on nodes to view music information, and explore related music pieces along the edges. This mode helps users intuitively understand the learning progression relationships and technical connections between music pieces. For learning sequence recommendations, the display rendering unit displays the music in the sequence sequentially, marking the position and learning stage of each piece in the sequence, allowing users to clearly understand the planning of the learning path.

[0162] The feedback collection unit gathers user feedback data, including ratings, learning completion rates, and learning time for recommended pieces. Rating feedback is obtained through explicit user ratings, typically using a five-star or ten-point scale. Users can rate a piece after completing it, expressing their level of enjoyment or learning experience. Learning completion feedback is inferred from user learning behavior, such as whether the user has completely learned all chapters of the sheet music, their progress in watching instructional videos, and the number of practice sessions. Learning time feedback records the total time spent by the user on each piece, including time spent viewing the sheet music, watching instructional videos, and actual practice (if the user uses a compatible smart piano or MIDI input device, the system can record the actual performance time). The feedback collection unit stores this data in the user's historical learning record and feeds it back to the hybrid recommendation module for updating the collaborative filtering matrix. It also feeds it back to the learning sequence generation module for dynamic adjustments to the ability assessment, forming a closed-loop optimization mechanism for the recommendation system.

[0163] The user interaction module enhances the user experience through a user-friendly interface and diverse display methods, while providing data support for continuous system optimization through effective feedback collection.

[0164] The data synchronization module is responsible for synchronizing data between the cloud database and the user's local application, ensuring consistent data access across different devices and supporting offline use. This module includes a difference detection unit, an incremental transmission unit, and a conflict resolution unit.

[0165] The difference detection unit compares the data versions in the cloud database and the local cache to identify the data items that need to be synchronized. Preferably, data synchronization adopts a timestamp-based version control mechanism. Each data record maintains a last modified timestamp. The difference detection unit determines which data has changed by comparing the timestamps in the cloud and locally. For data whose cloud timestamp is newer than the local timestamp, it indicates that the cloud data has been updated and needs to be downloaded locally; for data whose local timestamp is newer than the cloud timestamp, it indicates that the local data has been modified and needs to be uploaded to the cloud. The difference detection unit generates a synchronization task list, where each task specifies the data item to be synchronized, the synchronization direction (upload or download), and the synchronization priority.

[0166] The incremental transmission unit performs the actual data transmission, transmitting only the changed data content, thus reducing network bandwidth consumption. For metadata synchronization, the incremental transmission unit transmits changed data records; for file synchronization, it uses file segmentation and difference comparison technology to divide the file into several data blocks, calculate the hash value of each data block, and compare the hash values ​​between the local and cloud servers to transmit only the changed data blocks, avoiding the duplicate transmission of unchanged data. This technology is particularly suitable for synchronizing large files. For example, when a segment of an instructional video is modified, only the modified data block needs to be transmitted, rather than the entire video file, significantly reducing the amount of data transmitted. The incremental transmission unit supports breakpoint resumption; if a network interruption causes a transmission failure, the next transmission can continue from the point of interruption without retransmitting the completed portion.

[0167] The conflict resolution unit handles conflicts arising when the same data item is modified simultaneously on both the cloud and local machines. Preferably, the conflict resolution strategy includes two modes: timestamp priority and user confirmation. The timestamp priority mode employs a last-write-wins strategy, retaining the most recent version and discarding older versions. This mode is suitable for scenarios where data loss is unlikely, such as synchronizing user preferences. The user confirmation mode, upon detecting a conflict, displays the conflict details to the user, allowing the user to choose which version to keep or manually merge the two versions. This mode is suitable for synchronizing important data, such as user-created personal study plans or notes. The conflict resolution unit records the occurrence and resolution of conflicts for system quality monitoring and improvement.

[0168] The data synchronization module employs a strategy combining periodic background synchronization and triggered synchronization. Periodic background synchronization checks for data discrepancies and performs synchronization at regular intervals (e.g., every 30 minutes or at fixed times each day) when the user's device is connected to the internet, ensuring eventual data consistency. Triggered synchronization executes immediately when the user performs a specific action (e.g., explicitly clicking the synchronization button, opening the application, or completing a music lesson), meeting the user's need for real-time data. Through the combination of these two synchronization strategies, the data synchronization module ensures data consistency while also optimizing network resource utilization and user experience.

[0169] To verify the effectiveness of the technical solution of this invention, a systematic comparative experiment was conducted. The experiment was carried out on a music library containing 5,000 piano pieces and a real dataset of 1,000 piano learners, comparing the recommendation effect of the system of this invention with that of the prior art CN 105718566 A.

[0170] Regarding recommendation accuracy, the Hit Rate@10 was used as the evaluation metric, representing the percentage of the top 10 recommended tracks actually learned by the user. Experimental results show that the Hit Rate@10 of this invention's system reaches 76.3%, while the existing technology's Hit Rate@10 is 63.8%, representing an improvement of 12.5 percentage points. Particularly in the cold start scenario (where the user's historical learning records are less than 5 tracks), the Hit Rate@10 of this invention's system is 58.2%, while the existing technology is only 38.5%, representing an improvement of 19.7 percentage points, fully demonstrating the advantages of hybrid recommendation algorithms in solving the cold start problem.

[0171] Regarding instructional suitability, ten piano teachers were invited to rate the rationality of the recommended instructional results, with scores ranging from 1 to 5. The average instructional suitability score for this invention's system was 4.5, while the average score for existing technologies was 3.7, representing an improvement of 0.8 points, or 21.6%. Teacher evaluations indicated that the learning sequence recommended by this invention's system has a clearer difficulty gradient and a more reasonable technical progression, conforming to the gradual progression principle of piano instruction.

[0172] Regarding learning efficiency, the learning progress of 100 learners using the system of this invention and 100 learners using existing technology systems was tracked over the same period (6 months). Standardized piano skill tests were used to assess the learners' ability improvement. The experimental results showed that learners using the system of this invention achieved an average skill improvement of 18.7%, while those using existing technology systems achieved 15.3%, resulting in a 22.2% increase in learning efficiency.

[0173] Regarding system response speed, the average response time of recommendation requests was tested. The system of this invention, through topology optimization of the music knowledge graph and shortest path pre-calculation, controls the recommendation response time to within 150 milliseconds, meeting the requirements of real-time recommendations. In contrast, existing technologies require path searching within the song distance network for each recommendation, resulting in an average response time of 400 milliseconds. The system of this invention improves response speed by 62.5%.

[0174] The experimental results above show that the piano music library cloud management and intelligent recommendation system provided by the present invention is significantly superior to the existing technology in multiple dimensions such as recommendation accuracy, teaching adaptability, learning efficiency and system response speed, verifying the effectiveness and advancement of the technical solution.

[0175] The piano music library cloud management and intelligent recommendation system provided by this invention achieves in-depth modeling of the professional characteristics of piano pieces through a music knowledge graph construction module, effectively solves the cold start problem and improves recommendation accuracy through a hybrid recommendation module, provides learners with a systematic and progressive learning path planning through a learning sequence generation module, ensures the legal use of digital resources through a copyright management module, and achieves efficient data synchronization between the cloud and local systems through a data synchronization module. The entire system adopts a modular design, with each module working collaboratively to achieve intelligent organization and personalized delivery of piano teaching resources, providing a complete technical solution for the digital transformation of piano education.

[0176] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, which may include: read-only memory, random access memory, disk or optical disk, etc.

[0177] The above provides a detailed description of the piano music library cloud management and intelligent recommendation system provided by the embodiments of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A cloud-based piano music library management and intelligent recommendation system, characterized in that: The system includes: The music knowledge graph construction module is used to build a structured music knowledge graph that includes difficulty levels, technical characteristics, and musical styles. The hybrid recommendation module combines music content features and user collaborative filtering information to generate a personalized recommendation list; The learning sequence generation module is used to plan a progressive learning path in the music knowledge graph based on the learner's ability assessment results. The cloud database module is used to store music metadata, digital scores, and teaching video resources; The copyright management module is used to control access permissions for music resources and track usage records; The user interaction module is used to receive user operation instructions, display recommendation results, and collect learning feedback data. The data synchronization module is used to achieve incremental synchronization between cloud databases and local applications.

2. The piano music library cloud management and intelligent recommendation system according to claim 1, characterized in that, The music knowledge graph construction module includes: The feature extraction unit is used to extract multi-dimensional features such as difficulty level, technical type, musical style, composer, and creation period from the music metadata. The graph structure generation unit is used to represent music as nodes in a graph, and constructs directed edges based on difficulty gradient, technical similarity and style relevance to form a music knowledge graph; Topology optimization units are used to adjust the graph structure to optimize the reachability of learning paths by utilizing connectivity analysis and path optimization methods in graph theory.

3. The piano music library cloud management and intelligent recommendation system according to claim 2, characterized in that, The graph structure generation unit constructs three types of edges: Difficulty progression edges indicate a learning order between two pieces of music, with edge weights calculated based on the difficulty gradient and technical compatibility. Technical similarity edges represent the similarity between two pieces of music in terms of technical training objectives. The edge weights are calculated based on the cosine similarity of the technical feature vectors and the difficulty matching. Style association edges indicate the similarity of two pieces of music in terms of musical style, and are used to recommend tracks with similar or diverse styles.

4. The piano music library cloud management and intelligent recommendation system according to claim 1, characterized in that, The hybrid recommendation module includes: The content feature matching unit is used to calculate the matching degree between candidate music pieces and learners' ability features; The collaborative filtering unit is used to predict the current learner's preference for candidate music pieces based on the historical learning records of similar learners; The hybrid fusion unit is used to fuse content matching scores and collaborative filtering scores through adaptive weights to generate a comprehensive recommendation score.

5. The piano music library cloud management and intelligent recommendation system according to claim 4, characterized in that, The adaptive weights of the hybrid fusion unit are dynamically adjusted based on the amount of historical interaction data of the learner. When there is less historical interaction data, the weight of the content matching score is higher. As the interaction data increases, the weight of the collaborative filtering score gradually increases.

6. The piano music library cloud management and intelligent recommendation system according to claim 1, characterized in that, The learning sequence generation module includes: The competency assessment unit is used to quantitatively assess a learner's current competency level based on their historical learning records and test results. The path search unit is used to search for learning paths that satisfy the increasing difficulty constraint in the music knowledge graph, starting from the music node corresponding to the learner's current ability. The sequence optimization unit is used to optimize the generated learning path, balancing path length, difficulty gradient, and content diversity.

7. The piano music library cloud management and intelligent recommendation system according to claim 6, characterized in that, The sequence optimization unit evaluates candidate learning sequences using a comprehensive scoring function, which includes three dimensions: difficulty smoothness score, content diversity score, and interest score.

8. The piano music library cloud management and intelligent recommendation system according to claim 1, characterized in that, The cloud database module includes: Metadata storage unit, used to store structured metadata of the music, including attributes such as difficulty, style, and technical characteristics; The file storage unit is used to store image files of digital sheet music and multimedia files of teaching videos; The index building unit is used to create a multidimensional index for music metadata, supporting fast retrieval.

9. The piano music library cloud management and intelligent recommendation system according to claim 1, characterized in that, The copyright management module includes: The access control unit is used to set access permissions for different music resources based on the user's subscription status and payment history; A tracking unit is used to record the time, frequency, and duration of a user's access to music resources; The watermark embedding unit is used to embed invisible digital watermarks in digital sheet music and instructional videos to identify the user.

10. The piano music library cloud management and intelligent recommendation system according to claim 1, characterized in that, The data synchronization module includes: The difference detection unit is used to compare the data versions between the cloud database and the local cache to identify the data items that need to be synchronized. Incremental transmission units are used to transmit only the data content that has changed, thereby reducing network bandwidth consumption. The conflict resolution unit is used to handle conflicts when the same data item is modified simultaneously in the cloud and locally, and merges them using a strategy that prioritizes timestamps or requires user confirmation.

Images