Improved Neural Collaborative Filtering Algorithm for Personalized Training Plan Recommendations

To address the data sparsity and cold start issues in collaborative filtering recommendation, the existing models mainly rely on user behavior data, ignoring the project content characteristics and user personal attributes. This study applies the improved neural collaborative filtering (I-NCF) algorithm to process high-dimensional sparse data through nonlinear feature combination to achieve more precise personalized training plan recommendations for football players. First, the athlete’s personal information and training plan data are collected and converted into low-dimensional vectors using the Embedding layer. Then, residual connections and attention mechanisms are applied in I-NCF to adaptively capture feature combinations. Different features are gradually fused through the hidden layer. Batch normalization is used to accelerate training convergence, and dropout is added to avoid overfitting. Finally, stochastic gradient descent (SGD) is utilized to adjust the network weights, and cross-entropy is utilized as the loss function to optimize the model hyperparameters based on grid search. The study shows that I-NCF performs well in recommending coordination and balance training plans, with a mean absolute error (MAE) of only 0.068; the average Precision@k in nine football training programs reaches 0.817, with a standard deviation of only 0.036; the speed of athletes increases by 12.80%, and the endurance increases by 15.40% in the fourth week of training; the training plan under the improved algorithm can significantly improve the physical fitness of players, improve training efficiency and achieve precise personalized recommendations.