Collaborative filtering is a popular recommendation algorithm that uses similarity metrics like Cosine, Pearson, and Jaccard to predict user preferences. However, traditional similarity metrics are not dynamic and do not adapt to changes in user rating density, which can limit the accuracy of predictions in sparse settings. This paper introduces an adaptive dynamic similarity strategy that adjusts similarity weights based on user profile attributes. The proposed strategy is tested as a standalone memory-based algorithm and as a combination of different machine learning regression models, such as linear regression, random forest, k-nearest neighbors, gradient boosting and support vector regression. To evaluate predictive accuracy, the MovieLens-100K and MovieLens-1M datasets are used, and the MAE and RMSE are employed as evaluation metrics. Results show that the cosine similarity metric is still a strong competitor among traditional algorithms, while the incorporation of the adaptive similarity idea through support vector regression learning leads to a significant decrease in the prediction error, with support vector regression performing best. The results clearly indicate that the integration of adaptive similarity modeling and machine learning can significantly increase the accuracy of the collaborative filtering approach.