A music recommendation method based on individual and group multi-view interest modeling
By constructing user-user and music-music similarity graphs, combining individual and group interest collaborative modeling, and using graph convolutional models to extract multi-layer features, the problem of inaccurate interest expression in traditional music recommendation is solved, achieving more accurate and personalized music recommendation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- 内蒙航天动力机械测试所
- Filing Date
- 2022-07-08
- Publication Date
- 2026-05-12
AI Technical Summary
Traditional music recommendation methods suffer from sparse historical interaction data, resulting in inaccurate expression of user interests, making it difficult for existing technologies to provide efficient personalized music recommendations.
By constructing user-user similarity interest graphs and music-music similarity attribute graphs, and employing collaborative modeling of individual and group interests, we utilize graph convolutional models to extract multi-layered user and music features, and combine them with deep graph propagation models for interest mining to achieve personalized music recommendations.
It alleviates the problem of data sparsity, improves the accuracy of music recommendations, and enables more comprehensive and personalized user profiles and music feature descriptions, thereby enhancing the accuracy of recommendations.
Smart Images
Figure CN115098780B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of music recommendation technology, and in particular to a music recommendation method based on multi-perspective interest modeling of individuals and groups. Background Technology
[0002] Music brings rich spiritual enjoyment, but the development of the music industry has led to an overload of musical works, making it difficult for users to efficiently find the music they want to listen to. To address this issue, more and more music software has emerged online, giving rise to music recommendation technology. This technology can provide users with personalized music playlists based on their music playback history, meeting their individual needs.
[0003] Current traditional music recommendation methods primarily rely on historical interaction data to obtain user interests and preferences, using methods such as collaborative filtering to generate interest-based recommendations. However, historical interaction records are often extremely sparse, leading to inaccurate learned interest representations. Therefore, learning rich, multi-perspective interests using limited sparse interactions to model more accurate user and music feature representations is particularly important for music recommendation technology. Summary of the Invention
[0004] Therefore, the purpose of this invention is to provide a music recommendation method based on multi-perspective interest modeling of individuals and groups. It utilizes limited sparse interactions to construct user-user similarity interest graphs and music-music similarity attribute graphs, alleviating the data sparsity problem. By employing collaborative modeling of individual and group interests to model user preferences and music attributes, it more comprehensively describes user profiles and music features, improving the accuracy of music recommendation technology. Furthermore, it uses a graph convolution model to obtain multi-layered user and music feature representations, mining deep user interests and deep music attributes to achieve personalized music recommendation based on multi-perspective interest modeling of individuals and groups.
[0005] The present invention solves the above-mentioned technical problems through the following technical means:
[0006] A music recommendation method based on multi-perspective interest modeling of individuals and groups, with the following specific steps:
[0007] S1: Collect the user's historical music playback records as raw data, and number the users and music in the raw data; map the users and music into the feature vector space, and obtain the initial feature representation of the user and the initial feature representation of the music through Gaussian initialization;
[0008] S2: Construct a user-music interaction graph based on the user's historical music playback records. This matrix reflects the individual user's interests and the individual attributes of the music.
[0009] S3: Construct a user-user similarity interest relationship graph based on shared music playback records. This matrix reflects the group similarity interests of users. Construct a music-music similarity attribute relationship graph based on shared listening user records. This matrix reflects the group similarity attributes of music.
[0010] S4: By combining the interaction graph and similarity graph reflecting the interests and preferences of individual and group users, a collaborative relationship graph of individual and group interests of users is obtained; by combining the interaction graph and similarity graph reflecting the attributes of individual and group music, a collaborative relationship graph of individual and group attributes of music is obtained.
[0011] S5: On the relationship graph of individual and group interest collaboration, use the depth graph propagation model to extract the deep interest features of users and the deep attribute features of music, and obtain multi-layer user feature vectors and music feature vectors containing multi-perspective interests of individuals and groups.
[0012] S6: Pool the depth map to obtain multi-layer user feature vectors and multi-layer music feature vectors, respectively, to obtain the final user feature vectors and final music feature vectors for individual and group multi-perspective interests;
[0013] S7: Calculate the similarity between the obtained user feature vectors and music feature vectors that contain multiple perspectives of individual and group interests. The similarity level represents the degree of preference. Sort each user’s preference for all the music to be recommended, select the top K music according to the size of the metric, and generate Top-K music recommendations for the user.
[0014] Furthermore, the user's historical music playback records are collected as raw data. The users and music in the raw data are then numbered, specifically by assigning a unique ID to each user and music. Gaussian initialization is then used to initialize the feature representations of the users and music, resulting in initial feature vectors for each.
[0015] Furthermore, the user-music interaction graph is derived from the interaction between users and music. The points on the graph represent all users and all music, and the edges on the graph represent the observable user-music interaction records in the original data.
[0016] Furthermore, the user-user similarity interest relationship graph originates from the similarity interest relationship between users, with the points on the graph representing all user nodes and the edges representing the user-user similarity relationships with shared interactive music that can be observed in the original data; the music-music similarity attribute relationship graph originates from the similarity attribute relationship between music, with the points on the graph representing all music nodes and the edges representing the music-music similarity relationships with shared interactive users that can be observed in the original data.
[0017] Furthermore, the user's individual and group interest collaboration graph is a combination of the user-music interaction graph and the user-user similar interest graph, which is combined by matrix addition; the music's individual and group attribute collaboration graph is a combination of the user-music interaction graph and the music-music similar attribute graph, which is combined by matrix addition.
[0018] Furthermore, a deep graph propagation model is used to learn user feature vectors and music feature vectors of multi-layered individual and group multi-perspective interests. Specifically, the deep graph propagation model is a graph convolutional depth model; a self-loop is added to adjust the proportion of interest information at different depths in the graph convolutional depth model.
[0019] Furthermore, the pooling method is a mean operation, yielding the final user feature vector and the final music feature vector that incorporate the collaborative interests of individuals and groups. The loss function in the deep model training process is the BPR (Bayesian Pairwise Loss) function, and the optimizer is the stochastic gradient descent optimization algorithm.
[0020] Furthermore, the spatial distance similarity between the user vector and the candidate music vector is calculated using the inner product; the similarity metrics are then sorted to obtain the Top-K music recommendations for the user.
[0021] The beneficial effects of this invention are:
[0022] This invention utilizes limited sparse interactions to construct user-user similarity interest graphs and music-music similarity attribute graphs. The user-user similarity interest graphs are used to describe the group interests of users, and the music-music similarity attribute graphs are used to describe the group attributes of music, thus alleviating the data sparsity problem.
[0023] This invention employs a collaborative modeling approach combining individual and group interests to model user preferences and music attributes, thereby providing a more comprehensive description of user profiles and music characteristics and improving the accuracy of music recommendation technology.
[0024] By using graph convolution models to obtain multi-layered user and music feature representations, we can mine deep user interests and deep music attributes. We can also use efficient inner product operations to calculate the similarity between user vectors and music vectors, thereby realizing personalized music recommendations based on multi-perspective interest modeling for individuals and groups. Attached Figure Description
[0025] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:
[0026] Figure 1 : Overall block diagram of the music recommendation method based on multi-perspective interest modeling of individuals and groups in this invention;
[0027] Figure 2 : A schematic diagram for constructing a graph of the collaborative relationships between individual and group attributes;
[0028] Figure 3 : Schematic diagram of multi-perspective interest modeling for individuals and groups;
[0029] Figure 4 : Schematic diagram of individual and group interest pooling and interest prediction. Detailed Implementation
[0030] To make the objectives, features, and advantages of the technical solution proposed in this invention more apparent and understandable, the embodiments of the technical solution proposed in this 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 proposed technical solution, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.
[0031] like Figure 1 As shown, the music recommendation method of the present invention based on multi-perspective interest modeling of individuals and groups specifically includes:
[0032] S1: Collect the user's historical music playback records as raw data, and number the users and music in the raw data; map the users and music into the feature vector space, and obtain the initial feature representation of the user and the initial feature representation of the music through Gaussian initialization;
[0033] S2: Construct a user-music interaction graph based on the user's historical music playback records. This matrix reflects the individual user's interests and the individual attributes of the music.
[0034] S3: Construct a user-user similarity interest relationship graph based on shared music playback records. This matrix reflects the group similarity interests of users. Construct a music-music similarity attribute relationship graph based on shared listening user records. This matrix reflects the group similarity attributes of music.
[0035] S4: By combining the interaction graph and similarity graph reflecting the interests and preferences of individual and group users, a collaborative relationship graph of individual and group interests of users is obtained; by combining the interaction graph and similarity graph reflecting the attributes of individual and group music, a collaborative relationship graph of individual and group attributes of music is obtained.
[0036] S5: On the relationship graph of individual and group interest collaboration, use the depth graph propagation model to extract the deep interest features of users and the deep attribute features of music, and obtain multi-layer user feature vectors and music feature vectors containing multi-perspective interests of individuals and groups.
[0037] S6: Pool the depth map to obtain multi-layer user feature vectors and multi-layer music feature vectors, respectively, to obtain the final user feature vectors and final music feature vectors for individual and group multi-perspective interests;
[0038] S7: Calculate the similarity between the obtained user feature vectors and music feature vectors that contain multiple perspectives of individual and group interests. The similarity level represents the degree of preference. Sort each user’s preference for all the music to be recommended, select the top K music according to the size of the metric, and generate Top-K music recommendations for the user.
[0039] Figure 2 This diagram illustrates the construction of the individual and group attribute collaboration relationship graph in this method. The specific steps include:
[0040] S1: Collect users' historical music playback records as raw data. Number the users and music tracks in the raw data, assigning each a unique ID. Use Gaussian initialization to define the feature representations of users {u|u∈U} and music tracks {i|i∈I}, obtaining initial feature vectors for users and music tracks respectively. and express.
[0041] S2: After obtaining the renumbered users and music from step S1, construct the user-music interaction graph R. ui ={r ui Given the graph |u∈U,i∈I}, the points represent all users and all music tracks, and the edges represent observable user-music interaction records from the original data. If user-music interaction is observed, then r ui =1, if no user-music interaction is observed, then r ui =0.
[0042] S3: Obtain the user-music interaction relationship R in step S2. ui Then, construct a user-user similarity interest relationship graph R. uu The points in the graph represent all user nodes, and the edges represent the observable user-user similarity relationships with shared interactive music in the original data.
[0043] R uu =R ui ·R ui T (1)
[0044] Where R ui T For R ui The transposition of reflects the music-user interaction relationship.
[0045] Construct a similarity attribute relationship graph between music-music pairs using R. ii The points in the graph represent all music nodes, and the edges represent the music-music similarity relationships observed in the original data among users with common interactions.
[0046] R ii =R ui T ·R ui (2)
[0047] Figure 3 This diagram illustrates the multi-perspective interest modeling of individuals and groups in this method. The specific steps include:
[0048] S4: By using the interaction graph R that reflects the interests and preferences of individual and group users. ui and similar interest graph R uu By combining these elements, we obtain the graph R representing the collaborative relationships between individual and group interests of users. u ,Right now
[0049] R u =R ui ·R uu (3)
[0050] By using the interaction diagram R that reflects the attributes of individual musicians and groups ui Combined with similar interest maps R ii The graph R represents the synergistic relationship between individual and group attributes of music. i ,Right now
[0051] R i =R ui T +R ii (4)
[0052] Figure 4 This diagram illustrates the individual and group interest pooling and interest prediction methods used in this approach. The specific steps include:
[0053] S5: In step S4, the individual and group interest collaboration relationship graphs R of the users were obtained respectively. u Diagram R showing the synergistic relationship between individual and group attributes in music. i The deep graph propagation model is used to learn user feature vectors E representing multi-layered individual and group interests from multiple perspectives. u l+1 and music feature vector E i l+1 The depth graph propagation model described therein is a graph convolutional depth model, with a self-loop added to adjust the proportion of interest information at different depths in the graph convolutional depth model.
[0054] E ul+1 =D u -1 / 2 R ui D i -1 / 2 ·E i l +D u -1 R uu ·E u l +α·E u l (5)
[0055] E i l+1 =D i -1 / 2 R ui T D u -1 / 2 ·E u l +D i -1 R ii ·E i l +α·E i l (6)
[0056] Where l∈[0,L], L is the number of layers in the depth map propagation model, and in this invention, L=3, D u This is a diagonal matrix describing the number of user interactions, where the values on the diagonal are the number of user interactions with music; D i The diagonal matrix describes the number of music interactions, with the values on the diagonal representing the number of interactions between the music and the user; α is a hyperparameter that adjusts the self-looping ratio, and in this invention, α = 0.1.
[0057] S6: After obtaining the multi-layered individual and group multi-perspective interest user features and music features from step S5, respectively, the multi-layered user feature vectors and multi-layered music feature vectors obtained from the depth maps are pooled to obtain the final user feature vectors and final music feature vectors for individual and group multi-perspective interests, i.e.
[0058]
[0059]
[0060] The loss function is the Bayesian pairwise loss (BPR) function, and the optimizer is the stochastic gradient descent algorithm. In the loss function, positive examples represent music that interacts with the user, and negative examples represent music for which no interaction was observed. The stochastic gradient descent algorithm is used to optimize the Bayesian pairwise loss.
[0061] S7: Obtain the final user feature vector and the final music feature vector E from step S6. u and E i The spatial distance similarity between user vectors and candidate music vectors is calculated using the inner product, i.e.
[0062] Y ui =E u T ·E i (9)
[0063] Where Y ui E represents the predicted score for user u and music i. u T Represents vector E u After transposing, the similarity metrics are sorted to obtain the user's Top-K music recommendations.
Claims
1. A music recommendation method based on multi-perspective interest modeling of individuals and groups, characterized in that, The specific steps are as follows: S1: Collect the user's historical music playback records as raw data, number the users and music in the raw data, map the users and music into the feature vector space, and obtain the initial feature representation of the user and the initial feature representation of the music through Gaussian initialization; S2: Construct a user-music interaction matrix R based on the user's historical music playback records. ui This matrix reflects individual user interests and individual attributes of music; S3: Construct a user-user similarity interest relationship graph based on shared music playback records, and a user-user similarity interest relationship matrix R. uu To reflect the similar interests of user groups, a music-music similarity attribute relationship graph is constructed based on the listening records of users who listen to the same music. The music-music similarity attribute relationship matrix R ii Reflecting the group similarity attribute of music; S4: By using the user-music interaction relationship matrix R that reflects individual user interests ui And the similarity interest relationship matrix R that reflects the preferences of user groups uu By combining these, we obtain the individual and group interest synergy matrix R of users. u By using the interaction matrix R that reflects the individual attributes of music ui T And the similarity attribute relationship matrix R reflecting the attributes of music groups ii By combining these, we obtain the synergistic relationship matrix R of individual and group attributes in music. i ; S5: On the graph of the relationship between individual and group interests, using a depth graph propagation model in R... u Extracting deep user features on R, and simultaneously in R i The deep attribute features of music are extracted to obtain multi-layered user feature vectors and music feature vectors containing the interests of individuals and groups from multiple perspectives; S6: Using the mean pooling method, the user feature vectors E obtained in S5 for each layer are processed. u l and the music feature vectors E of each layer i l Perform aggregation operations separately to obtain the final user feature vector E that incorporates interest information at different depths. u and the final music feature vector E i ; S7: The final user feature vector E obtained from S6, which contains multi-perspective interests of individuals and groups, is... u and the final music feature vector E i Similarity is calculated, with the degree of similarity representing the level of preference. Each user's preference for all the music to be recommended is ranked, and the top K music is selected according to the magnitude of the metric, generating Top-K music recommendations for the user.
2. The music recommendation method based on multi-perspective interest modeling of individuals and groups according to claim 1, characterized in that: Collect users' historical music playback records as raw data, and assign numbers to the users and music in the raw data. The specific steps are as follows: S11: Assign a unique number to users and music, and set a unique ID for each user and music; S12: Use Gaussian initialization to obtain the initial feature vectors of users and music.
3. The music recommendation method based on multi-perspective interest modeling of individuals and groups according to claim 1, characterized in that: The user-music interaction matrix is derived from the interaction between users and music. The points on the graph represent all users and all music, and the edges on the graph represent the observable user-music interaction records in the original data.
4. The music recommendation method based on multi-perspective interest modeling of individuals and groups according to claim 1, characterized in that: The user-user similarity interest relationship matrix is derived from the similarity interest relationship between users. The points on the graph are all user nodes, and the edges on the graph are the user-user similarity relationships with shared interactive music that can be observed in the original data. The music-music similarity attribute relationship graph is derived from the similarity attribute relationship between music. The points on the graph are all music nodes, and the edges on the graph are the music-music similarity relationships with shared interactive users that can be observed in the original data.
5. The music recommendation method based on multi-perspective interest modeling of individuals and groups according to claim 1, characterized in that: The individual and group interest synergy matrix is a combination of the user-music interaction matrix and the user-user similar interest matrix, which is combined by matrix addition. The individual and group attribute synergy matrix is a combination of the user-music interaction matrix and the music-music similar attribute matrix, which is combined by matrix addition.
6. The music recommendation method based on multi-perspective interest modeling of individuals and groups according to claim 1, characterized in that, The deep graph propagation model is used to learn user feature vectors and music feature vectors representing multi-layered individual and group interests from multiple perspectives. The specific operation is as follows: S51: The depth graph propagation model described therein is a graph convolution depth model; S52: Add a self-loop to adjust the proportion of interest information at different depths in the graph convolution depth model.
7. The music recommendation method based on multi-perspective interest modeling of individuals and groups according to claim 1, characterized in that: The pooling method is mean operation, which yields the final user feature vector and the final music feature vector containing the collaborative interests of individuals and groups. The loss function in the deep model training process is the Bayesian pairwise loss function, and the optimizer is the stochastic gradient descent optimization algorithm.
8. The music recommendation method based on multi-perspective interest modeling of individuals and groups according to claim 1, characterized in that: S71: Calculate the spatial distance similarity between user vectors and candidate music vectors using the inner product; S72: Sort the similarity metrics to obtain the user's Top-K music recommendations.