Predicting danceability and song ratings using deep learning and auditory features

Data collection

The dataset used in this project originates from Kaggle and contains approximately 19,000 songs with 15 features. Each observation represents an individual song. The majority of these features are numerical values that quantify various musical attributes such as loudness, instrumentalness, danceability, and liveness. Some columns consist of string-based data, a few of which were dropped, while others were converted from numeric to categorical values to better represent their information. The dataset includes two target columns: song_popularity, which assigns a popularity score from 0 to 100 for the popularity prediction task, and danceability for the danceability prediction task. The key features of the dataset are summarized in Table 1.

To better illustrate the distribution of features, Figs. 1 and 2 provide graphical representations of categorical and numerical features, respectively. Figure 3 visualizes box plots of song_popularity and danceability in the dataset. These visualizations offer insights into the structure and variability of the dataset.

Data preprocessing

To ensure the dataset’s compatibility and enhance the performance of the predictive model, several preprocessing steps were applied. The dataset consists of approximately 19,000 songs with 15 features, where most attributes are numerical, quantifying musical characteristics such as loudness, instrumentalness, danceability, and liveness. Some features were categorical, requiring encoding to facilitate their use in machine learning algorithms.

First, the dataset was loaded from a CSV file using Pandas. Features were classified into categorical (audio_mode, key, time_signature) and numerical types, with song_name excluded due to its irrelevance to the prediction task. Label encoding was applied to categorical features, converting them into numerical representations while preserving their distinctiveness. Subsequently, numerical features were standardized using the StandardScaler to ensure uniform scaling, which stabilizes the learning process. For the regression task, the label, which originally ranged from 0 to 100, was normalized to the range [0,1]. This transformation ensures that all values lie within a fixed, interpretable scale, preventing potential numerical instability in gradient-based optimization. Additionally, normalizing the target variable can help models converge faster by reducing the impact of different magnitudes between input features and the label. Following preprocessing, the dataset was randomly split into training, validation, and test sets in an 80-10-10 ratio. The test set was separated first, followed by an additional split of the training set to create the validation data. This approach ensures effective model evaluation and minimizes the risk of overfitting. The statistical information of the data in each of the training, validation, and test sets is summarized in the Table 2. These preprocessing steps collectively optimize the dataset, enabling the development of robust predictive models for estimating song popularity based on musical attributes.

Selection method

The selection of modelling techniques for this study was driven by the nature of the input data and the dual prediction task involving both danceability and song rating. The dataset contains heterogeneous features, including categorical metadata (e.g., genre, artist) and numerical auditory descriptors (e.g., tempo, loudness, spectral features), each requiring specialised handling. To effectively model the temporal and contextual patterns in the categorical features, we selected a BiLSTM network, which is well-suited for capturing sequential dependencies in structured inputs. For the numerical auditory features, we employed a Residual Network (ResNet), chosen for its ability to extract deep hierarchical representations while preserving gradient flow through residual connections.

Given the complementary nature of these feature types, we adopted a cross-attention mechanism to integrate the BiLSTM and ResNet outputs. This allows the model to dynamically learn complex interdependencies between the two modalities, enhancing its capacity to understand how auditory signals and metadata jointly influence a song’s perceived danceability and popularity. The selection of this hybrid framework was also informed by preliminary experiments, where simpler concatenation approaches yielded lower performance and failed to fully capture cross-modal relationships. Therefore, the final architecture was chosen based on both theoretical suitability and empirical validation.

Model architecture

Building on the selection rationale outlined above, we designed a hybrid deep learning architecture that leverages the strengths of both sequence modelling and hierarchical feature extraction. As illustrated in Fig. 4, the proposed model consists of two parallel processing streams tailored to the distinct modalities in the dataset.

Categorical features, such as genre or artist, are first embedded and passed through a BiLSTM network to capture temporal and contextual dependencies across sequences. In parallel, numerical auditory features, including tempo, spectral characteristics, and rhythmic patterns, are processed by a ResNet module to learn deep, multi-level representations that retain important low- and high-level acoustic information.

The outputs from both branches are then integrated using a cross-attention mechanism, which enables the model to dynamically align and weigh relevant information across the two modalities. This fusion strategy allows for the discovery of intricate relationships between auditory signals and categorical context. The combined representation is subsequently passed through a fully connected layer, which generates the final predictions for both danceability and overall song rating.

This architecture is designed to exploit the complementary strengths of BiLSTM and ResNet while enabling cross-modal interaction through attention, making it well-suited for modelling the complex dynamics of music perception and listener preferences.

Model configuration and training strategy

To support replicability and provide transparency in architectural design choices, we detail both the model configuration and training strategy employed in this study.

Computing resource

All experiments were conducted on a high-performance computing environment equipped with an Intel Xeon processor, 32 GB of RAM, and an NVIDIA RTX 3070 GPU with 12 GB of VRAM. The models were implemented using Python with the PyTorch framework. Data preprocessing and analysis were performed using Pandas and NumPy, while visualization was conducted with Matplotlib and Seaborn. The deep learning training process leveraged CUDA optimizations to enhance computational efficiency. Hyperparameter tuning was carried out to ensure optimal model performance. The entire pipeline was executed on Ubuntu 20.04 to provide a stable and efficient computational environment.

Evaluation metrics

To assess the performance of our model, we employ four widely used evaluation metrics: mean squared error (MSE), the coefficient of determination ( $R^2$), mean absolute error (MAE), and root mean square deviation (RMSE). These metrics provide a robust and comprehensive analysis of the model’s predictive capabilities.

• Mean squared error: measures the average squared difference between actual and predicted values, indicating the overall prediction error:

  (5) $$\mathrm{M}\mathrm{S}\mathrm{E}=\frac{1}{n}\sum _{i=1}^n{(y_i-{\hat{y}}_i)}^2.$$

• Coefficient of determination: represents the proportion of variance in the dependent variable explained by the model, with higher values indicating better performance:

  (6) $$R^2=1-\frac{{\sum }_{i=1}^n{(y_i-{\hat{y}}_i)}^2}{{\sum }_{i=1}^n{(y_i-\overline{y})}^2}.$$

• Mean absolute error: computes the average absolute difference between actual and predicted values, offering an interpretable error measure:

  (7) $$\mathrm{M}\mathrm{A}\mathrm{E}=\frac{1}{n}\sum _{i=1}^n|y_i-{\hat{y}}_i|.$$

• Root mean square deviation: calculates the square root of MSE, providing a more intuitive measure of error by maintaining the same unit as the target variable:

  (8) $$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{\frac{1}{n}\sum _{i=1}^n{(y_i-{\hat{y}}_i)}^2}.$$

Evaluation strategy

Ablation study

To further investigate the importance of individual components within our proposed architecture, we conducted a series of ablation experiments. These experiments were designed to systematically remove or replace key elements of the model, thereby isolating their effects on the prediction performance. The following ablation variants were implemented and evaluated:

• Without ResNet: the ResNet module responsible for processing numerical features was removed. Numerical inputs were instead passed through a single fully connected layer without residual connections, allowing us to evaluate the role of deep residual learning in capturing complex numerical patterns.

• Without BiLSTM: the BiLSTM block used for capturing contextual dependencies in categorical features was excluded. Categorical data were encoded and passed through a shallow dense layer, testing the importance of sequential modeling in this context.

• Replacing Cross-Attention with Concatenation: instead of using a cross-attention mechanism to fuse the BiLSTM and ResNet outputs, we concatenated both feature vectors and passed them directly to the final prediction layer. This setup explores whether attention-based integration is superior to simple feature merging.

• Replacing BiLSTM with GRU (Cho et al., 2014): to assess the sensitivity of the model to the choice of sequence encoder, we replaced the BiLSTM with a Gated Recurrent Unit (GRU). This variant maintains the same embedding and attention pipeline, providing insight into alternative temporal modeling strategies.

• Replacing BiLSTM with Transformer Encoder (Vaswani et al., 2017): in this configuration, we replaced the BiLSTM with a Transformer encoder to leverage its ability to model long-range dependencies and capture global context through self-attention mechanisms. Additionally, the Transformer enables more efficient parallel processing, leading to faster training and improved performance on large datasets.

The ablation results in Table 4 highlight the importance of each architectural component in our proposed model. Removing the ResNet block leads to a noticeable degradation in both tasks, particularly in the popularity prediction ( $R^2$ drops from 0.3698 to 0.2810) and danceability prediction ( $R^2$ drops from 0.6443 to 0.5976), indicating its critical role in capturing informative visual features. Similarly, excluding the BiLSTM module causes a reduction in performance across all metrics, reaffirming its effectiveness in modeling temporal dependencies. The cross-attention mechanism also proves essential, as its removal leads to a consistent performance drop in both tasks. For instance, $R^2$ on the danceability task falls from 0.6443 to 0.6147, demonstrating the advantage of guided interaction between multimodal features over naive concatenation. Additionally, replacing BiLSTM with alternative sequence encoders, both GRU and Transformer variants yield slightly inferior results. Additionally, replacing BiLSTM with an alternative sequence encoder, bib8, yields a slightly better result. The GRU-based model achieves an $R^2$ of 0.3421 (popularity) and 0.6289 (danceability). The GRU-based model achieves an $R^2$ of 0.3421 (popularity) and 0.6289 (danceability), while the Transformer-based version records 0.3367 and 0.6240, respectively. These results suggest that although the architecture is relatively robust to changes in the sequence encoder, BiLSTM remains the most effective choice for our tasks. Overall, the full model outperforms all ablated versions, validating the design choices made in our framework.

Comparative study

The results presented in Table 5 confirm the effectiveness of the proposed model in both prediction tasks. Our model achieves the best performance across all evaluation metrics, with the lowest MAE (0.1228 and 0.0667), MSE (0.0297 and 0.0089), and RMSE (0.1723 and 0.0942) for the popularity and danceability tasks, respectively. Additionally, it attains the highest $R^2$ scores—0.3842 for popularity and 0.6443 for danceability—indicating that it captures a substantially greater portion of variance compared to all baseline models. Among the traditional machine learning methods, XGBoost yields the best performance, achieving an MAE of 0.1613 (popularity) and 0.0906 (danceability), with corresponding $R^2$ scores of 0.1407 and 0.4831. This highlights the effectiveness of gradient boosting in handling structured features. Random Forest and AdaBoost perform comparably, benefiting from ensemble learning, though their predictive power remains clearly lower than the proposed model. On the other hand, Decision Tree and $k$-Nearest Neighbors exhibit weaker performance, likely due to their limited generalization capabilities and tendency to overfit or underfit complex patterns in the data.

In summary, the experimental results validate the effectiveness of our approach, demonstrating a substantial improvement over traditional methods. The performance of the proposed model can be attributed to its ability to capture complex relationships within the dataset, possibly leveraging deep learning or more advanced ensemble techniques.

The performance comparison in Table 6 highlights the effectiveness of our proposed model over a range of state-of-the-art deep learning approaches tailored for tabular data. Across both prediction tasks, our model consistently achieves the best results in all evaluation metrics. Specifically, it records the lowest MAE (0.1228 for popularity and 0.0667 for danceability), MSE (0.0297 and 0.0089), and RMSE (0.1723 and 0.0942), demonstrating both high accuracy and reduced prediction error. Moreover, it attains the highest $R^2$ scores (0.3842 and 0.6443), indicating a substantially better ability to explain the variance in the target variables compared to all baseline deep learning models. Among the compared methods, GNM shows the strongest baseline performance, achieving an $R^2$ of 0.2219 (popularity) and 0.5737 (danceability). However, our model outperforms GNM by a notable margin, particularly in the popularity task, where the $R^2$ improvement exceeds 16 percentage points. Additionally, both TabTransformer and FT-Transformer deliver competitive results, reflecting the strength of attention mechanisms in capturing complex feature interactions. FT-Transformer improves upon TabTransformer with a lower MAE (0.1428 vs. 0.1541) and higher $R^2$ (0.2450 vs. 0.2130) in the popularity task, but still falls short of our model’s performance across all metrics. These gains suggest that our model captures more complex and relevant relationships within the data, potentially due to improved architectural design or better multimodal feature integration. While MLP, DCN, and TabNet provide moderate results, they consistently lag behind in both tasks, underscoring the limitations of standard or shallow deep learning architectures when applied to this problem domain. In contrast, our model’s strong performance across the board indicates enhanced representational capacity and generalization ability.

The experimental findings validate the effectiveness of our approach, demonstrating a significant improvement over current deep learning models for tabular data. The consistent reduction in prediction errors and the notable gains in explanatory power position our model as a promising solution for music-related predictive tasks. Future work may explore adaptations of this architecture for other music attributes or expand its use to real-world recommendation systems.