A music recommendation method based on knowledge graph multi-task feature learning

Abstract

The present invention discloses a music recommendation method based on multi-task feature learning of knowledge graph, comprising: step 1, in the knowledge graph construction stage, based on the Satori library and user-music interaction records, the feature vectors of entities and relationships obtained by training are mapped to low-dimensional vectors to construct the required knowledge graph; step 2, in the model training stage, the feature representations of users and their neighborhood nodes on the knowledge graph are fused through neighborhood aggregation operations, the local neighbor structure is captured and stored in each entity, and the user preference features are obtained; step 3, in the recommendation stage, the potential features of music are predicted through the CNN regression model, and the matching degree between the user and the music is calculated in combination with the user preference features, and a recommendation list that the final user may be interested in is generated. A comparative experiment is conducted on the Last.FM dataset, and the results show that compared with the traditional recommendation method and the knowledge graph-based method, the algorithm can effectively mine the potential relationship between users and music and improve the quality of music recommendation.

Description

Technical Field

[0001] The present invention belongs to the field of computer applications, and specifically relates to a music recommendation method based on knowledge graph multi-task feature learning. Background Art

[0002] In recent years, with the rapid development of cloud computing, big data, the Internet of Things and other technologies, the emergence of various applications in the Internet space has triggered an explosive growth in data size. In this new digital information age, massive data resources are provided to users, causing them great trouble. The emergence of recommendation systems has effectively alleviated this problem.

[0003] In the recommendation task of multimedia products, music recommendation has gradually received widespread attention. Currently popular music recommendation algorithms are mainly divided into content-based recommendation algorithms, collaborative filtering-based recommendation algorithms and hybrid recommendation algorithms. Among them, the music recommendation method based on collaborative filtering is widely used. It can automatically filter and select target users based on similar users and further explore the potential interests of users. However, when faced with the sparsity problem, the collaborative filtering algorithm is also difficult to guarantee the quality of music recommendations.

[0004] In recent years, the cold start and sparsity problems in music recommendation scenarios have attracted the attention of more and more scholars. Zhang et al. proposed a personalized recommendation algorithm to solve the problem of single recommendation method in multimedia products. Zheng et al. proposed a dynamic music recommendation framework based on tags, which uses the characteristics of available semantic tags of music tracks to supplement the highly sparse user-item interaction matrix, greatly simplifying the difficulty of recommending tracks for new users. Hariri et al. first classified the music using the user's social tags before recommending, and then classified and recommended the user's situation based on the relevant information in the user's playback record, achieving good recommendation results.

[0005] However, recommendations based only on audio, tags or categories will lose a lot of music feature information, thus failing to provide users with good personalized services. In actual recommendation scenarios, the music cold start problem is ubiquitous. For new users and music, they cannot be recommended due to the lack of relevant feedback data, which greatly affects the user's listening experience.

[0006] Therefore, it is difficult to build a good feature system using a single category of data, and it is of great significance to integrate multi-source auxiliary information for music recommendation. Summary of the invention

[0007] The present invention is made to solve the above-mentioned problems, and its purpose is to provide a music recommendation method based on knowledge graph multi-task feature learning.

[0008] The present invention provides a music recommendation method based on knowledge graph multi-task feature learning, which has the following characteristics and includes the following steps: Step 1, in the knowledge graph construction stage, based on the Satori library and user-music interaction records, the feature vectors of entities and relationships obtained by training are mapped to low-dimensional vectors to construct the required knowledge graph; Step 2, in the model training stage, the feature representations of users and their neighborhood nodes on the knowledge graph are fused through neighborhood aggregation operations, the local neighbor structure is captured and stored in each entity, and user preference features are obtained; Step 3, in the recommendation stage, the potential features of music are predicted through a CNN regression model, and the matching degree between users and music is calculated in combination with user preference features, so as to generate a recommendation list that the final user may be interested in.

[0009] In the music recommendation method based on knowledge graph multi-task feature learning provided by the present invention, it can also have the following characteristics: wherein, step 1 specifically includes the following sub-steps: step 1-1, in the entity linking stage, mapping the users, music and their attribute values ​​in the data set to the corresponding entities; step 1-2, in the relationship extraction stage, treating the implicit interactive behaviors such as listening, purchasing, and clicking between users and music as relationships; step 1-3, in the knowledge storage stage, mapping the feature vectors of the trained entities and relationships into low-dimensional vectors to construct the required knowledge graph.

[0010] The music recommendation method based on knowledge graph multi-task feature learning provided by the present invention may also have the following features: In step 2, all user-item pairs (u, v) are traversed, and the receptive field M of item v is calculated in a layer-by-layer iterative manner and the model parameters are initialized. By calculating the neighborhood linear combination of V, the user relationship score is used as a filter to select a set of items that the user may be interested in. Finally, the two aggregation functions, Mean aggregator and Concat aggregator, are used to aggregate the entity representation v and its neighborhood representation into a vector.

[0011] The music recommendation method based on knowledge graph multi-task feature learning provided by the present invention may also have such features, and also include: step 4, the detection stage, in which data from different fields are collected and brought into the model, personalized recommendations are performed, and the optimal detection results are obtained.

[0012] Functions and Effects of the Invention

[0013] According to the music recommendation method based on knowledge graph multi-task feature learning involved in the present invention, the specific steps are as follows: Step 1, in the knowledge graph construction stage, based on the Satori library and user-music interaction records, the feature vectors of the trained entities and relationships are mapped to low-dimensional vectors to construct the required knowledge graph; Step 2, in the model training stage, the feature representations of the user and its neighborhood nodes on the knowledge graph are fused through the neighborhood aggregation operation, the local neighbor structure is captured and stored in each entity, and the user preference features are obtained; Step 3, in the recommendation stage, the potential features of the music are predicted through the CNN regression model, and the matching degree between the user and the music is calculated in combination with the user preference features to generate a recommendation list that the final user may be interested in.

[0014] Therefore, compared with the conventional benchmark model, the music recommendation method based on knowledge graph multi-task feature learning of the present invention can effectively mine the potential relationship between users and music and improve the quality of music recommendation through the KGHR algorithm proposed in the present invention. The model first combines the Satori library and user-music interaction records to construct a music knowledge graph, maps the feature vectors of entities and relationships trained from the knowledge graph into low-dimensional vectors, and introduces the recommendation system, which intuitively shows the potential connection between users and music. At the same time, the joint training method is adopted to closely link the recommendation task and the knowledge graph training task, and achieves a good recommendation effect; secondly, the algorithm obtains the user's preference for the music knowledge graph through neighborhood aggregation operations, calculates the scores of users and relationships, and thus more accurately recommends songs that may be of interest to users. It can be seen that obtaining the user's personalized preferences and integrating the semantic information of the knowledge graph can help improve the performance of the recommendation system.

[0015] In addition, the core idea of ​​the hybrid music recommendation model KGHR (Hybrid Recommendation based on Knowledge Graph) proposed in the present invention is to aggregate and merge neighborhood information in a biased manner when calculating the representation of a given entity in the knowledge graph. On the one hand, this design enables users to more conveniently obtain local information and structure of neighboring nodes through neighborhood aggregation operations; on the other hand, the neighborhood is weighted according to the relationship between entities and the ratings of specific users. It reflects both the semantic information between entities in the knowledge graph and the user's personalized interest in the relationship, greatly improving the accuracy and interpretability of the recommendation and making the recommendation results traceable.

[0016] Finally, the present invention utilizes the integration of multi-source heterogeneous data from the knowledge graph, maps the feature vectors of entities and relationships trained from the knowledge graph into low-dimensional vectors, and introduces a recommendation system to fully utilize the knowledge information of the knowledge graph. In addition, the algorithm does not require manual design of meta-paths. The model expands KGHR from one layer to multiple layers through neighborhood aggregation operations, thereby mining the user's potential interests more broadly and deeply. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a model diagram of a music recommendation method based on knowledge graph multi-task feature learning in an embodiment of the present invention;

[0018] Figure 2 This is a code operation flow chart of the music recommendation method based on knowledge graph multi-task feature learning in an embodiment of the present invention. DETAILED DESCRIPTION

[0019] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the following embodiments and the accompanying drawings specifically illustrate a music recommendation method based on knowledge graph multi-task feature learning of the present invention.

[0020] In this embodiment, a music recommendation method based on knowledge graph multi-task feature learning is provided.

[0021] Figure 1 It is a model diagram of the music recommendation method based on knowledge graph multi-task feature learning in an embodiment of the present invention.

[0022] Figure 2 This is a code operation flow chart of the music recommendation method based on knowledge graph multi-task feature learning in an embodiment of the present invention.

[0023] like Figure 1-2 As shown, the music recommendation method based on knowledge graph multi-task feature learning involved in this embodiment includes the following steps:

[0024] Step 1, knowledge graph construction phase: Based on the Satori library and user-music interaction records, the feature vectors of entities and relationships obtained through training are mapped into low-dimensional vectors to construct the required knowledge graph. The specific implementation steps are:

[0025] Step 1-1, in the entity linking stage, users, music and their attribute value mappings in the dataset are linked to the corresponding entities.

[0026] Step 1-2, in the relationship extraction stage, implicit interactive behaviors such as listening, purchasing, and clicking between users and music are regarded as relationships.

[0027] Steps 1-3, in the knowledge storage stage, map the feature vectors of entities and relationships obtained through training into low-dimensional vectors to construct the required knowledge graph.

[0028] Step 2, training phase: mainly to obtain data and train the model. In the modeling phase, this embodiment uses neighborhood aggregation operations to fuse the feature representations of users and their neighborhood nodes on the knowledge graph, thereby obtaining the user's potential remote preferences. Traverse all user-item pairs (u, v), calculate the receptive field M of item v in a layer-by-layer iterative manner and initialize the model parameters, calculate the neighborhood linear combination of V, use the user relationship score as a filter, and select the set of items that the user may be interested in. Finally, the two aggregation functions, Mean aggregator and Concat aggregator, are used to aggregate the entity representation v and its neighborhood representation into a vector, which greatly improves the recommendation performance.

[0029] Step 3, in the recommendation module, this embodiment uses the CNN regression model to predict the potential features of music, and calculates the matching degree between the user and the music in combination with the user preference features, and finally generates a recommendation list that the user may be interested in.

[0030] Step 4, detection phase: mainly to collect data, bring it into the model, and obtain the best detection result. This embodiment uses the musician listening information of 2,000 users from the Last.FM online music system as a data set for music recommendation. After filtering the data set, the experiment divides the Last.FM data set into: training set, test set and validation set in a ratio of 8:1:1. The model parameters are selected on the validation set, and AUC, ACC and F1 are used to evaluate the CTR prediction performance of the model. Two commonly used evaluation indicators, accuracy (Precision@20) and recall (Recall@20), are selected to measure the recommendation effect of the model.

[0031] In this example, the KGHR model is compared with the collaborative knowledge base embedding model (CKE), the knowledge graph-aware neural network model (KGNN-LS), the ripple set recommendation model (RippleNet) and the multi-task enhanced recommendation model (MKR). The results show that the proposed method has an improvement of 7.8% and 2% in Recall@20 and Precision@20, respectively, which verifies the effectiveness of the proposed method.

[0032] Functions and Effects of the Embodiments

[0033] According to the music recommendation method based on knowledge graph multi-task feature learning involved in this embodiment, the specific steps are as follows: Step 1, in the knowledge graph construction stage, based on the Satori library and user-music interaction records, the feature vectors of the trained entities and relationships are mapped to low-dimensional vectors to construct the required knowledge graph; Step 2, in the model training stage, the feature representations of the user and its neighborhood nodes on the knowledge graph are fused through the neighborhood aggregation operation, the local neighbor structure is captured and stored in each entity, and the user preference features are obtained; Step 3, in the recommendation stage, the potential features of the music are predicted through the CNN regression model, and the matching degree between the user and the music is calculated in combination with the user preference features to generate a recommendation list that the final user may be interested in.

[0034] Therefore, compared with the conventional benchmark model, the music recommendation method based on knowledge graph multi-task feature learning in this embodiment can effectively mine the potential relationship between users and music and improve the quality of music recommendation. The model first combines the Satori library and user-music interaction records to construct a music knowledge graph, maps the feature vectors of entities and relationships trained from the knowledge graph into low-dimensional vectors, and introduces the recommendation system, which intuitively shows the potential connection between users and music. At the same time, the joint training method is adopted to closely link the recommendation task and the knowledge graph training task, and achieves good recommendation results; secondly, the algorithm obtains the user's preference for the music knowledge graph through neighborhood aggregation operations, calculates the scores of users and relationships, and thus more accurately recommends songs that may be of interest to users. It can be seen that obtaining the user's personalized preferences and integrating the semantic information of the knowledge graph can help improve the performance of the recommendation system.

[0035] In addition, the core idea of ​​the hybrid music recommendation model KGHR (Hybrid Recommendation based on Knowledge Graph) proposed in this embodiment is to aggregate and merge neighborhood information in a biased manner when calculating the representation of a given entity in the knowledge graph. On the one hand, this design enables users to more conveniently obtain local information and structure of neighboring nodes through neighborhood aggregation operations; on the other hand, the neighborhood is weighted according to the relationship between entities and the ratings of specific users. It reflects both the semantic information between entities in the knowledge graph and the user's personalized interest in the relationship, greatly improving the accuracy and interpretability of the recommendation and making the recommendation results traceable.

[0036] Finally, this embodiment utilizes the integration of multi-source heterogeneous data from the knowledge graph, maps the feature vectors of entities and relationships trained from the knowledge graph into low-dimensional vectors, and introduces a recommendation system to fully utilize the knowledge information of the knowledge graph. In addition, the algorithm does not require manual design of meta-paths. The model expands KGHR from one layer to multiple layers through neighborhood aggregation operations, thereby mining the user's potential interests more broadly and deeply.

[0037] The above-mentioned embodiments are preferred examples of the present invention and are not intended to limit the protection scope of the present invention.

Claims

1. A music recommendation method based on knowledge graph multi-task feature learning, characterized in that: The following steps are involved: Step 1: In the knowledge graph construction phase, based on the Satori library and user-music interaction records, the feature vectors of entities and relationships obtained through training are mapped into low-dimensional vectors to construct the required knowledge graph. Step 2: In the model training phase, the feature representations of the user and its neighboring nodes on the knowledge graph are fused through neighborhood aggregation operations, the local neighbor structure is captured and stored in each entity, the user preference features are obtained, all user-item pairs (u, v) are traversed, the receptive field M of item v is calculated in a layer-by-layer iterative manner and the model parameters are initialized. By calculating the neighborhood linear combination of V, the user relationship score is used as a filter to select the set of items that the user may be interested in. Finally, the entity representation v and its neighborhood representation are aggregated into a vector using the Mean aggregator and Concat aggregator aggregation functions; Step 3, in the recommendation stage, the potential features of the music are predicted through the CNN regression model, and the matching degree between the user and the music is calculated in combination with the user preference features to generate a recommendation list that the end user may be interested in.

2. The music recommendation method based on knowledge graph multi-task feature learning according to claim 1 is characterized in that: in, Step 1 specifically includes the following sub-steps: Step 1-1, in the entity linking stage, users, music and their attribute value mappings in the dataset are linked to corresponding entities; Step 1-2, in the relationship extraction stage, the implicit interaction behaviors such as listening, purchasing, clicking, etc. between the user and the music are regarded as relationships; Step 1-3, in the knowledge storage stage, the feature vectors of the entities and the relationships obtained through training are mapped into low-dimensional vectors to construct the required knowledge graph.

3. The music recommendation method based on knowledge graph multi-task feature learning according to claim 1 is characterized in that: Also includes: Step 4, detection phase, During the detection phase, data from different fields are collected, brought into the model, and personalized recommendations are made to obtain the best detection results.

Images