In this paper, we first constructed a knowledge tracking model and embedded a self-attention mechanism to optimise the model and overcome the problem of data sparsity. Subsequently, the structures of the input layer, embedding layer, and hidden layer are sequentially designed to build the knowledge tracking model. Using neurons, the learning matrix was passed into a linear transformation, and a linear model was employed to obtain the predicted values. Individual student feedback on learning progress is calculated using the EVA evaluation model, and the timeliness feedback for all teaching subjects is compiled into an ensemble that constitutes the overall immediate feedback. Stage feedback is established in the same way to constitute a feedback mechanism for music teaching, and a strategy for supporting feedback teaching decisions is proposed. The experimental samples are selected, the teaching practice environment is established, and the results of visual tracking of learning progress are analyzed. The mean values of the total music learning strategy scores of the experimental class and the control class were 26.4136 and 19.229, respectively. Further analyses showed that there was a significant difference between the experimental class and the control class under the 95% confidence intervals of the music learning dimensions, assuming that the ANOVA values were all 0.05. The number of irrelevant learning behaviours of the students in the experimental and control classes was 95 and 137, respectively, and overall it seems that feedback on students’ learning progress can reduce the number of invalid behaviours.