Predicting enjoyment as a number, not a category
The study asks whether EEG recorded while people listen to music carries enough information to estimate how much they enjoyed a song, and whether that estimate holds for listeners the model has never seen. Earlier music-EEG work, as the paper describes it, mostly treated emotion as a categorical label or used black-box models, and often split data at the trial level, which the authors argue inflates apparent generalization to new listeners.
The target here is a regression output: a continuous enjoyment score on a 1 to 5 Likert scale that each participant gave after hearing a song. The input is a multichannel EEG segment with C channels and T time samples, and the model learns a mapping from that segment to a predicted score. Ratings were normalized to [0, 1] during training and remapped to the 1 to 5 scale for evaluation, so all reported errors are in original scale points.
Holding out five listeners changes the benchmark
The data come from MUSIN-G, a public dataset collected at the Indian Institute of Technology Gandhinagar, containing 128-channel EEG from 20 participants who listened to 12 song excerpts of about two minutes each across multiple genres, with eyes closed to reduce ocular artifacts. After each song, participants rated overall enjoyment on the 1 to 5 scale.
The critical design choice is the participant-wise protocol: 15 listeners were used for training and validation, and 5 were held out entirely for testing. This prevents subject-level data leakage, so test performance reflects prediction for new listeners rather than for new segments of already-seen listeners. That choice is also why the reported correlations should be read as cross-listener generalization, not as within-listener decoding accuracy.
Song-level labels make every window equally enjoyable
Preprocessing follows a standard pipeline: average referencing, a 0.2 Hz high-pass filter, ICA-based artifact rejection with MARA via EEGLAB, downsampling to 250 Hz, interpolation of bad channels with spherical splines, and segmentation into fixed non-overlapping windows of 500 samples.
Because enjoyment was rated once per song, every window cut from that song inherits the same label. The authors call this a weak-label approach and note that it is useful for temporal modeling but introduces label noise, since enjoyment naturally fluctuates across a full song. This matters for interpretation: a model trained this way cannot learn moment-to-moment enjoyment changes, and the explanations describe which parts of a two-second window drive a song-level prediction, not which moments a listener actually enjoyed most.
Lowest error and best trend alignment go to different models
Six architectures were compared: EEGNet, ShallowConvNet, DeepConvNet, CNN-LSTM, TCN, and GNN-CNN. All were optimized with a combined loss weighting mean squared error at 0.6 and mean absolute error at 0.4, using Adam or AdamW with an initial learning rate of 3×10⁻⁴, cosine annealing, dropout of 0.5 to 0.6, and early stopping after 5 to 7 epochs without validation improvement.
On the held-out five listeners, GNN-CNN had the lowest mean absolute error at 0.97, followed by CNN-LSTM at 0.98 and DeepConvNet at 0.99. CNN-LSTM had the lowest RMSE at 1.18, which the paper reads as the best behavior against large errors. The TCN had the highest Pearson correlation at r = 0.21, described as the best trend alignment, while GNN-CNN had the highest tolerance accuracy at 58.3%. EEGNet and ShallowConvNet were clearly weaker, with MAE of 1.03 and 1.31 and correlations of −0.01 and −0.06.
The authors report that a statistical comparison of MAE between DeepConvNet and TCN showed no significant difference, which they interpret as spatial and temporal architectures capturing complementary aspects of enjoyment. The practical reading is that no single model simultaneously minimizes absolute error and maximizes correlation, so the choice of model depends on whether a use case cares about being close in scale or about ranking listeners' enjoyment correctly.
A correlation of 0.21 is a weak signal, not a decoder
The headline correlation deserves careful reading. An r of 0.21 means the model captures only a modest linear relationship between predicted and reported enjoyment across held-out listeners, even though its average absolute error is under one point on a five-point scale. Those two results are not contradictory: a model can sit near the middle of the rating range and achieve moderate absolute error while still tracking individual differences poorly.
This is why the paper's own framing separates absolute accuracy from trend alignment. The lowest-MAE models and the highest-correlation model are different, and the tolerance accuracy figures, which range from 35.0% to 58.3%, describe how often predictions fall within an accepted margin rather than how well they rank songs. Readers comparing these numbers should not treat a low MAE as evidence of strong enjoyment decoding, nor a 0.21 correlation as evidence that EEG carries no enjoyment information.
Explanations point to sustained segments, not isolated events
Three post-training explanation methods were applied to the TCN: Grad-CAM, input gradient sensitivity, and occlusion-based analysis. Gradient-based and occlusion-based explanations were averaged over all held-out windows to identify which time segments within each two-second window most influenced predictions.
The reported pattern is that relevance is highest at the start of the window and decays over time in the Grad-CAM visualization, and that temporal relevance is distributed across sustained EEG segments rather than isolated events. The authors interpret this as support for a context-driven view in which enjoyment is a sustained, evolving cognitive state. This is an interpretation of model behavior, not a neural mechanism proven by the experiment; the explanations describe where the TCN's predictions are sensitive, not which brain processes cause enjoyment.
A benchmark for weak cross-listener prediction
The study's durable contribution is methodological: a reproducible participant-wise benchmark for EEG-based enjoyment regression on a public dataset, with six architectures compared under one protocol and explanations reported alongside standard metrics. It shows that EEG signals support meaningful but weak cross-subject enjoyment prediction, and that explainability analysis adds information that MAE and correlation alone do not provide.
Its boundaries follow from the same design. The evidence covers 20 listeners, 12 excerpts of roughly two minutes, eyes-closed recording, and song-level labels inherited by all windows, so the results do not establish moment-by-moment enjoyment tracking or transfer to other datasets, genres, or recording setups. The paper is an arXiv preprint, and its code is available from the corresponding author on request rather than as a released repository.
The judgment worth carrying forward is a comparison, not a single score: if the goal is to be close in scale to reported enjoyment, GNN-CNN, CNN-LSTM, and DeepConvNet are the relevant choices at MAE 0.97 to 0.99; if the goal is to track relative enjoyment trends across new listeners, the TCN's r = 0.21 is the best available here, and it is still weak. Any application should be planned around that gap between absolute accuracy and trend alignment rather than around either metric alone.
Sources used in this answer
EEG-Based Prediction of Music Enjoyment Using Deep Learning with Explainable Temporal Analysis
Under a participant-wise split of MUSIN-G (15 listeners train/validation, 5 held out), six deep learning models predicted song-level enjoyment with MAE 0.97 to 1.31 on a 1 to 5 scale; GNN-CNN had the lowest MAE (0.97) and TCN the highest Pearson correlation (r = 0.21).