A personalized recommendation method based on auxiliary features
By using a dual-tower model and a rating representation method, the problem of insufficient feature modeling in existing music recommendation systems is solved, achieving more accurate user preference representation and personalized recommendation effects.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- UNIV OF ELECTRONICS SCI & TECH OF CHINA
- Filing Date
- 2023-08-30
- Publication Date
- 2026-05-19
AI Technical Summary
Existing personalized music recommendation systems fail to fully utilize users' music playback history, resulting in insufficient feature modeling and affecting recommendation effectiveness.
A dual-tower model is adopted, including a convolutional neural network and two parallel neural networks. The quality features of user and song comment texts are extracted through user-side networks and song-side networks, respectively, and then cross-fused in the hidden space. Combined with user playback duration and song price features, a rating representation model is established and the loss function is optimized.
It improves the accuracy and interpretability of personalized recommendations, enhances the richness of feature information and the credibility of recommendations, and allows users to better understand the recommendation results.
Smart Images
Figure CN117235302B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of recommender systems, and more specifically to a personalized recommendation method based on auxiliary features. Background Technology
[0002] With the continuous development of information technology, the internet has become an important channel for people to obtain information. A large number of popular musical works are rapidly shared and disseminated through the internet, gaining attention and love from listeners. To solve the problem of users struggling to select music that matches their preferences from a vast selection, online music platforms often proactively provide personalized recommendations based on users' music playback data, displaying these recommendations on a personalized recommendation page.
[0003] Most existing personalized music recommendation systems resemble e-commerce recommendation systems, using deep neural networks (DNNs), RNNs, and Transformers as the basic structure for feature modeling and extraction. These models use user music playback history information, such as song ID, song category, and listening duration, to obtain corresponding feature embedding vectors through an embedding layer. These vectors are then trained on the corresponding neural network model to obtain a trained parameter model for subsequent prediction. The prediction layer calculates a final score for the target song based on a set objective function. This score reflects the similarity between the target song and the music in the user's playback history. Finally, the most suitable songs are recommended to the user based on this score—a Top-K recommendation. However, while existing personalized music recommendation methods utilize user music playback history information, they do not fully leverage it. Furthermore, they fail to effectively model and represent these features during feature cross-referencing, which weakens the effectiveness of personalized recommendations to some extent.
[0004] Therefore, it is necessary to design a new music recommendation method that fully utilizes users' music playback history information and achieves further efficient modeling to improve the effectiveness of personalized recommendations. Summary of the Invention
[0005] The purpose of this invention is to propose a personalized recommendation method based on auxiliary features to overcome the shortcomings of existing methods that fail to fully utilize users' music playback records for effective feature modeling and representation, thereby obtaining more accurate user vector representations and enabling the recommendation system to better provide personalized music recommendation services.
[0006] To achieve the above objectives, the present invention adopts the following technical solution:
[0007] A personalized recommendation method based on auxiliary features includes the following steps:
[0008] S1. Acquire music data and preprocess it to obtain a preprocessed dataset. The music recommendation data includes user features, song features, and user-song interaction features. User features include user ID, comment text, user rating, purchase price, and collection status. Song features include song ID, author ID, song category, song duration, and song price.
[0009] S2. Construct a dual-tower model to predict user preferences for the song, which will then be used for personalized online music recommendations. The dual-tower model consists of a convolutional neural network, two parallel neural networks, and a hidden layer space.
[0010] The convolutional neural network extracts text features from the comment text in the preprocessed dataset. The text features are the vector representing the user's original comment information on the song. If the user comments on the target song, the vector is output as the vector representing the user's original comment information on the song. Otherwise, the vector representing the user's original comment information on the song is calculated based on the song's comment feature vector.
[0011] The two parallel neural networks are a user-side network and a song-side network. The user-side network extracts the quality features of user comment text based on user features and text features; the song-side network extracts the quality features of user comment text on the target song based on song features and text features.
[0012] The quality features of user comment text and the quality features of user comment text on the target song are input into the hidden layer space, cross-fused to obtain high-order combined features, and then output to obtain the user's preference information for the song.
[0013] S3. Based on the user preference information for the target song obtained in S2, construct a user rating representation model for the song; establish a loss function based on squared error for the root rating representation model;
[0014] S4. Using the loss function established in S3, the gradient descent algorithm is used to optimize the loss function and train the dual-tower model. The trained dual-tower model is used as the standard model for subsequent online personalized music recommendations.
[0015] Furthermore, the user-side network has the same structure as the song-side network, both including an embedding layer, a self-attention block extraction module, and a comment attention network connected in sequence.
[0016] Furthermore, in the convolutional neural network, the detailed steps for extracting text features when the user has not commented on the target song are as follows:
[0017] a1. Filter out the text features of all user comments on the target song i to obtain the comment list r.i ;
[0018] a2. Aggregate the text features of the k most recent comments from the current user u to obtain the user's recent feature preferences r. uk , to be used to represent feature preferences over a recent period;
[0019] a3, in the comment list r i The text features of all comments related to user u and target song i are filtered out, and the recent feature preference feature r of user u is calculated. uk The similarity between user u and all comment text features in target song i is calculated, and the t feature text vectors with the highest similarity are obtained. The similarity calculation method in this embodiment adopts the Jaccard similarity calculation formula, as shown below:
[0020]
[0021] Where A and B represent two text feature vectors participating in the similarity calculation, the numerator represents the common set of elements of the two text features, and the denominator represents the union of the elements of the two text features;
[0022] a4. Calculate the average of the t most similar feature text vectors, and use the result as the text representation vector O of the item review. out This completes the text feature acquisition.
[0023] Furthermore, the method for obtaining comment text quality features using the two parallel neural networks includes the following sub-steps:
[0024] The method for obtaining the quality features of user-side online comment text is as follows:
[0025] b1. Input user features other than comment text into the embedding layer for embedding learning to obtain user embedding feature vectors;
[0026] b2. Input the user-embedded feature vector into the self-attention block, and use the self-attention mechanism to calculate the attention coefficients of various features from the user-embedded feature vector;
[0027] b3. Using the attention coefficients of various features obtained in b2 and the original comment information vector of the user on the song as initial features, the comment vector attention coefficient is calculated using the song comment attention network to obtain the quality representation result of the user comment vector, which is then output as the quality feature of the comment text.
[0028] The method for obtaining the quality features of the song-side network comment text is the same as that used for the side network. The only difference is that the song-side network obtains the quality features of the comment text of users commenting on the target song based on the song features.
[0029] Furthermore, the hidden space of S2 employs element-level dot product operations to fuse the quality features of user comment texts and the quality features of user comment texts on the target song, thereby obtaining user preference information for the song.
[0030] Furthermore, the scoring representation model formula in S3 is as follows:
[0031]
[0032] Where W0 represents the weight matrix of the prediction layer, b u1 b represents the normalized feature vector of music playback duration; u2 Let represent the equivalent music price perception feature vector, where ω1 and ω2 are two learnable weight coefficients used to adjust the user's perception of b. u1 b u2 The degree of preference between two vectors, where μ represents the global bias term;
[0033] The loss function is expressed as follows:
[0034]
[0035] Where S represents the sample instances in the training set, and R... u，i This represents real user ratings for the target item.
[0036] Furthermore, the data preprocessing method in S1 is as follows:
[0037] S1.1 Filter out special songs to ensure the equality of song author characteristics; special songs refer to songs with vague author information or songs with independent users as the author; vague author information means the author is unknown, a group of artists, etc., and independent users means the author is displayed as a non-singer.
[0038] S1.2 Cleaning comment text. The method for cleaning comment text is as follows:
[0039] First, filter out all non-English language comment text. Then, remove meaningless comment information. Finally, based on the activity level of the interaction, filter out individual user comments with fewer than a preset number of comments and comments for a single song with fewer than a preset number of comments.
[0040] S1.3. Convert the format of the data obtained in S1.2 to obtain the preprocessed dataset.
[0041] By adopting the above technical solution, the present invention has the following beneficial effects:
[0042] (1) This invention employs two parallel neural networks to construct a dual-tower model. Both parallel neural networks are based on dual self-attention networks. Based on the user's original comment information representation vector for the song, they respectively obtain the quality representation results of the user comment vector from the user features and obtain the quality representation results of the user's comment vector for the target song from the song features. This unique design enables parallel feature training with higher independence and efficiency, and avoids bias in modeling and learning due to early feature crossover causing confusion of higher-order features.
[0043] (2) This invention addresses the problem that the expression level of user interest is still lacking after the auxiliary features are preprocessed. It uses an embedding layer to obtain the embedding matrix of the input features and combines a self-attention mechanism to obtain the quality representation result of the comment vector, thus realizing high-order deep implicit feature extraction. On this basis, it introduces two features, song duration and equivalent price, to realize the combination of user interest vector representation and user personalized preference features, thereby enhancing the richness of feature information.
[0044] (3) Since the personalized recommendation method based on auxiliary features proposed in this invention is based on multiple features derived from the analysis results of user profiles, the interpretability of its personalized recommendation results is relatively good, and users have a higher degree of understanding and acceptance of the recommendation results, thus enhancing the interpretability of music recommendations. Attached Figure Description
[0045] Figure 1 This is a schematic diagram of the twin-tower model structure of the present invention;
[0046] Figure 2 This is a flowchart illustrating the implementation of the personalized recommendation method based on auxiliary features of this invention. Detailed Implementation
[0047] This invention aims to propose a method for personalized recommendation based on auxiliary feature information, addressing the shortcomings of current online platforms in personalized music recommendation, namely the lack of auxiliary feature information and the inefficiency in modeling user personalized preferences, thereby better achieving the task of personalized music recommendation. Its overall implementation process is as follows: Figure 2As shown, firstly, a music recommendation dataset is obtained from a music platform, and data preprocessing and format conversion are performed according to recommendation requirements to obtain sequence recommendation data that meets the requirements. The preprocessed data includes various feature data, which can be divided into user features, song features, and user-song interaction features. Next, a dual-tower model with two parallel neural network structures is established to model user-side features and song-side features respectively. After feature extraction by the two neural networks and a convolutional neural network, the features on both sides interact in a shared hidden space to achieve high-order feature combination. Then, the extracted features are decomposed according to preferences for comment text and other features, and a rating prediction model is established based on this. Simultaneously, to better model user personalized preference features, playback duration features and song price features are introduced as trainable parameters in the rating prediction model, and a squared loss error function is established. The target object for loss calculation is the rating prediction result and the actual rating result, and gradient descent is used for model training. The optimal model is obtained through several rounds of training. During online music recommendation testing, the loss function score of each song is calculated using the optimal dual-tower model. Finally, the K most relevant songs are selected as recommendation results and returned to the user, thus completing personalized music recommendation.
[0048] Example
[0049] The following section will use a typical online music platform as a recommendation scenario template to introduce and elaborate on a method for personalized music recommendation based on auxiliary feature information provided in this embodiment. This method includes the following steps: Figure 1 As shown;
[0050] S1: Obtain the music recommendation dataset and perform data preprocessing. The recommendation dataset used in this study is the QQ Music dataset, as shown in Table 1. This dataset contains user features, song features, and user-song interaction features. User features include user ID, comment text, user rating, purchase price, and collection status. Song features include song ID, author ID, song category, song duration, and song price.
[0051] The data preprocessing process includes three steps: filtering out special songs, cleaning comment text, and data format conversion.
[0052] S1.1 Filter out special songs to improve the quality of feature representation during personalized recommendations. The specific operation is as follows:
[0053] Songs with vague author information are filtered out to ensure the equality of the "author" characteristics of songs; special songs refer to songs with vague author information and the author being an independent user; vague author information refers to songs with unknown authors, multiple artists, etc.
[0054] In addition, considering the dissemination and popularity characteristics of songs, it is necessary to filter out songs whose authors are independent users, meaning the authors are not presented as singers.
[0055] S1.2 Cleaning Comment Text. User comments on target songs contain very important and rich information about user preferences, therefore, the comment text information needs to be strictly processed; otherwise, it will seriously affect the recommendation effect. The method for cleaning comment text is as follows:
[0056] First, filter out all non-English language comments to unify the comment language;
[0057] Then, meaningless comments are removed through manual review to ensure the quality of the comment text;
[0058] Finally, based on the principle of interaction activity, comments from individual users with fewer than a preset number of comments, and comments for a single song with fewer than a preset number of comments, are filtered out. In this embodiment, the number of comments from individual users and comments for a single song are both set to 3.
[0059] S1.3 Data Format Conversion. To ensure the data format meets the recommended requirements, it needs to be processed, including standardizing the time format and filling in short text. In this embodiment, the time format is uniformly converted to Unix time format, and short comment text is filled in using deep learning. The final preprocessed dataset D is obtained. It should be noted that in this embodiment, short text refers to text with tens to hundreds of characters, little content, and indistinct features.
[0060] The table below shows the auxiliary features involved in this recommendation.
[0061] Table 1 Classification of Auxiliary Features
[0062]
[0063]
[0064] S2: Establish a dual-tower model, inputting features, song features, and user-song interaction features into the dual-tower model; predict user preference information for the song based on the dual-tower model, for subsequent online personalized music recommendations.
[0065] The dual-tower model comprises a convolutional neural network (CNN), two parallel neural networks, and a hidden layer space. The CNN extracts text features representing the user's original comment information about the target song from the preprocessed dataset. If the user comments on the target song, the representation vector of that comment text is output as the text feature; otherwise, the text vector representing the comment is calculated based on the song's comment feature vector and output as the text feature.
[0066] The detailed process of extracting text features using a convolutional neural network is as follows:
[0067] The convolutional neural network in this embodiment includes an embedding layer, a convolutional layer, a pooling layer, and a fully connected layer. The pooling layer uses Max-Pooling pooling. The original comment information representation vector O corresponding to comment r can be obtained through text features. When a user makes a valid comment on the target song, the original comment information representation vector O is a non-zero vector; otherwise, it is a vector padded with zeros. To obtain valuable text feature information even in the absence of user comment features on the target song, this embodiment performs the following processing:
[0068] a1. Filter out the text features of all user comments on the target song i to obtain the comment list r. i ;
[0069] a2. Aggregate the text features of the k most recent comments from the current user u to obtain the user's recent feature preferences r. uk , to be used to represent feature preferences over a recent period;
[0070] a3, in the comment list r i The text features of all comments related to user u and target song i are filtered out, and the recent feature preference feature r of user u is calculated. uk The similarity between user u and all comment text features in target song i is calculated, and the t feature text vectors with the highest similarity are obtained. The similarity calculation method in this embodiment adopts the Jaccard similarity calculation formula, as shown below:
[0071]
[0072] Where A and B represent two text feature vectors participating in the similarity calculation, the numerator represents the common set of elements of the two text features, and the denominator represents the union of the elements of the two text features;
[0073] a4. Calculate the average of the t most similar feature text vectors, and use the result as the text representation vector O of the item review. out This completes the text feature acquisition.
[0074] The process of extracting quality features using two parallel neural networks is as follows:
[0075] The two parallel neural networks are a user-side network and a song-side network. The user-side network extracts the quality features of user comment text based on user features and text features; the song-side network extracts the quality features of user comment text on the target song based on song features and text features.
[0076] The user-side network and the song-side network have the same structure, both including an embedding layer, a self-attention block extraction module, and a comment attention network connected in sequence.
[0077] The method for obtaining the quality features of user-side online comment text is as follows:
[0078] b1. User features, excluding comment text, are input into the embedding layer for embedding learning to obtain user embedding feature vectors. In this embodiment, user features and song features are divided into ID features and other features. For other features, feature concatenation is performed before embedding so that features of different categories can be modeled in the same feature space. After being input into the embedding layer, both the user-side network embedding layer and the song-side network embedding layer will obtain two feature embedding matrices, namely the ID matrix and the attribute feature matrix.
[0079] b2. Input the user-embedded feature vector into the self-attention block, and use the self-attention mechanism to calculate the attention coefficients of various features from the user-embedded feature vector, thereby obtaining the weights of each feature;
[0080] b3. Using the attention coefficients of various features obtained in b2 and the original comment information vector of the user on the song as initial features, the comment vector attention coefficient is calculated using the song comment attention network to obtain the quality representation result of the user comment vector, which is then output as the quality feature of the comment text.
[0081] The method for obtaining the quality features of the song-side network comment text is the same as that used for the side network. The only difference is that the song-side network obtains the quality features of the comment text of users commenting on the target song based on the song features.
[0082] This embodiment uses a three-layer attention network stacked together to better extract user personalized feature preference information. The parameter configurations of the embedding layer network and the self-attention network are shown in Tables 2 and 3, respectively. The calculation formula for a single-layer attention network is as follows:
[0083]
[0084] Where Q, K, and V represent the input feature embedding matrix multiplied by three learnable training matrices W, respectively. Q WK W V The obtained parameter matrix, i represents the number of layers in the stacked attention network, S i This represents the output of the self-attention block in layer i.
[0085] Table 2. Configuration of Embedded Layer Network Parameters
[0086]
[0087] Table 3 Attention Network Parameter Configuration
[0088]
[0089] The text features from the convolutional neural network and the weights of each feature calculated by the self-attention block are input into the comment attention network. The user-side network extracts the quality features of the user's comment text; the song-side network extracts the quality features of the user's comment text on the target song.
[0090] Since the quality features of the comment text extracted by the user-side network and the song-side network in the dual-tower network model are independent and do not interact with each other, this embodiment introduces a shared hidden layer space to obtain deeper feature modeling representation capabilities. In this hidden layer space, element-wise dot product operations are used to cross-fuse the quality features of the user's comment text and the quality features of the user's comment text on the target song, thereby obtaining the user's preference information for the song. The representation of the feature cross-feature is shown below:
[0091] h0=(r u +X u )⊙(r i +Y i )
[0092] Here, h0 represents the output vector of the hidden layer space. To more accurately model user features, feature decomposition is used, decomposing the user and item feature preference representations based on review features and other attribute features. Specifically, r u and r i X represents the user's review preferences for items and the characteristics of item reviews, respectively. u and Y i These represent the user's preference for other features of the item and the preference for other features of the item, respectively.
[0093] S3. Based on the user preference information for the song obtained in S2, construct a user rating representation model for the song; establish a loss function based on squared error for the root rating representation model.
[0094] Rating representation model The formula is expressed as follows:
[0095]
[0096] Here, W0 represents the weight matrix of the prediction layer. To more accurately model user characteristics, two implicit preference features are introduced, namely b, based on the user's music playback duration and the price at which the user purchased music. u1 , represents the normalized feature vector representation of music playback duration; b u2 This represents the feature vector representation of the equivalent perceived price of music. ω1 and ω2 are two learnable weight coefficients used to adjust the magnitude of the two user preference vectors; μ represents the global bias term. The following focuses on b. u1 and b u2 The significance of this will be explained.
[0097] From the perspective of user behavior, the duration of a user's playback of a song significantly impacts preference modeling, directly reflecting the degree of user preference. However, considering the substantial differences in playback duration across different songs, normalization is necessary when incorporating this information. The calculation is shown below:
[0098]
[0099] Where T i This represents the total playback duration of song i in the user's playback history; t i This represents the duration of song i. It's easy to see that T... i >t i And the latter is a constant, therefore when T i The larger the value of b, the better. u1 The larger the value of b, the more it indirectly reflects the user's interest in the target song from the perspective of time experience. Similarly, in the loss function calculation process, it can be seen that b... u1 The value range is [0,1];
[0100] Considering price factors, a user's purchase of a song also reflects their preference for that song. For song i in the dataset, this paper converts the song's price into an equivalent price feature p based on its price characteristics. i Its definition expression is as follows:
[0101]
[0102] When the price of a song is 0, meaning it can be listened to without paying, p i The value is 1; when a song requires a membership (with a monthly fee of v and >1) to listen to, p i for When a song requires payment to purchase (assuming the song price is p), p i Let p be the value of music. Based on this, we can give the equivalent feature vector b of music price perception. u2 The expression is as follows:
[0103] b u2 =arctan(p i -0.9)
[0104] From the graph of the arctanx function, we can see that when the value of x is in the range [0.1, 2], the corresponding p i When the value is in the range [1, 2.9], the ratio between price factors and feature preferences fluctuates significantly. i Beyond this range, the function's fluctuation trend becomes relatively gentle.
[0105] According to the scoring prediction model Establish a loss function L based on the squared error. o Its formula is expressed as follows:
[0106]
[0107] Where S represents a sample instance in the training set; R u，i This represents real user ratings for the target item.
[0108] S4. Using the loss function established in S3, the gradient descent algorithm is used to optimize the loss function and train the dual-tower model. The trained dual-tower model is used as the standard model for subsequent online personalized music recommendations.
[0109] This embodiment uses PyTorch's Adam optimizer, with two exponential decay rates β1 and β2 of 0.9 and 0.999, respectively. The batch size is set to 1024, the learning rate to 0.01, and the number of epochs to 200. After repeated training, the model corresponding to the optimal loss is set as the standard model for the algorithm, used for subsequent online personalized music recommendations.
[0110] Although the present invention has been described herein with reference to embodiments thereof, the above embodiments are merely preferred embodiments of the present invention, and the implementation of the present invention is not limited to the above embodiments. It should be understood that those skilled in the art can design many other modifications and implementations, which will fall within the scope and spirit of the principles disclosed in this application.
Claims
1. A personalized recommendation method based on auxiliary features, characterized in that: Includes the following steps: S1. Acquire music data and preprocess it to obtain a preprocessed dataset. The music recommendation data includes user features, song features, and user-song interaction features. User features include user ID, comment text, user rating, purchase price, and collection status. Song features include song ID, author ID, song category, song duration, and song price. S2. Construct a dual-tower model to predict user preferences for the song, which will then be used for personalized online music recommendations. The dual-tower model consists of a convolutional neural network, two parallel neural networks, and a hidden layer space. The convolutional neural network extracts text features from the comment text in the preprocessed dataset. The text features are the vector representing the user's original comment information on the song. If the user comments on the target song, the vector is output as the vector representing the user's original comment information on the song. Otherwise, the vector representing the user's original comment information on the song is calculated based on the song's comment feature vector. The two parallel neural networks are a user-side network and a song-side network. The user-side network extracts quality features of user comment text based on user features and text features. The song-side network extracts the quality features of user comments on the target song based on song features and text features; The quality features of user comment text and the quality features of user comment text on the target song are input into the hidden layer space, cross-fused to obtain high-order combined features, and then output to obtain the user's preference information for the song. S3. Based on the user preference information for the target song obtained in S2, construct a user rating representation model for the song; establish a loss function based on squared error for the root rating representation model; S4. Using the loss function established in S3, the gradient descent algorithm is used to optimize the loss function and train the dual-tower model. The trained dual-tower model is used as the standard model for subsequent online personalized music recommendations.
2. The personalized recommendation method based on auxiliary features as described in claim 1, characterized in that: The user-side network and the song-side network have the same structure, both including an embedding layer, a self-attention block extraction module, and a comment attention network connected in sequence.
3. The personalized recommendation method based on auxiliary features as described in claim 1, characterized in that: In the convolutional neural network, the detailed steps for extracting text features when the user has not commented on the target song are as follows: a1. Filter all users' requests for the target song By analyzing the text features of the comments, a list of comments can be obtained. ; a2. Aggregate current users By analyzing the text features of the user's k most recent comments, we can obtain the user's recent feature preferences. , to be used to represent feature preferences over a recent period; a3. In the comment list Users were filtered out. With the target song Calculate user comment text features. Recent characteristics and preference characteristics With users With the target song The similarity of all comment text features is calculated, and the t feature text vectors with the highest similarity are obtained. The similarity is calculated using the Jaccard similarity formula, as shown below: in and Let represent the two text feature vectors involved in the similarity calculation, where the numerator represents the common set of elements of the two text features, and the denominator represents the union of the elements of the two text features; a4. Calculate the average of the t most similar feature text vectors, and use the result as the target song. The text representation vector O of the comment out This completes the text feature acquisition.
4. The personalized recommendation method based on auxiliary features as described in claim 2, characterized in that: The method for obtaining comment text quality features using the two parallel neural networks includes the following sub-steps: The method for obtaining the quality features of user-side online comment text is as follows: b1. Input user features other than comment text into the embedding layer for embedding learning to obtain user embedding feature vectors; b2. Input the user-embedded feature vector into the self-attention block, and use the self-attention mechanism to calculate the attention coefficients of various features from the user-embedded feature vector; b3. Using the attention coefficients of various features obtained in b2 and the original comment information vector of the user on the song as initial features, the comment vector attention coefficient is calculated using the song comment attention network to obtain the quality representation result of the user comment vector, which is then output as the quality feature of the comment text. The method for obtaining the quality features of the song-side network comment text is the same as that used for the side network. The only difference is that the song-side network obtains the quality features of the comment text of users commenting on the target song based on the song features.
5. The personalized recommendation method based on auxiliary features as described in claim 1, characterized in that: The hidden space of S2 employs element-level dot product operations to fuse the quality features of user comment texts and the quality features of user comment texts on the target song, thereby obtaining the user's preference information for the song.
6. The personalized recommendation method based on auxiliary features as described in claim 1, characterized in that: The formula for the scoring representation model in S3 is: in, This represents the weight matrix of the prediction layer. Represents the normalized feature vector of music playback duration; This represents the equivalent music price perception feature vector. and Two learnable weight coefficients are used to adjust the user's preference for... , The degree of preference between the two vectors Indicates the global bias term; The loss function is expressed as follows: in, For training set sample instances, This represents real user ratings for the target item.
7. The personalized recommendation method based on auxiliary features as described in claim 1, characterized in that: The data preprocessing method in S1 is as follows: S1.1 Filter out special songs to ensure the equality of song author characteristics; special songs refer to songs with vague author information or songs with independent users as the author; vague author information means the author is unknown or a mixed group of artists, and independent users means the author is displayed as a non-singer. S1.2 Cleaning comment text. The method for cleaning comment text is as follows: First, filter out all non-English language comment text. Then, remove meaningless comment information. Finally, based on the activity level of the interaction, filter out individual user comments with fewer than a preset number of comments and comments for a single song with fewer than a preset number of comments. S1.
3. Convert the format of the data obtained in S1.2 to obtain the preprocessed dataset.