Today, we trained a new model, now with about 60 movie genres, which lead to a much better accuracy with about 98% on the training set. This supports our experiences that models with more features are better capable to find patterns in the training data. However, to avoid over fitting, larger models also need more data. Thus, we plan to enhance our data set with more movies and the corresponding meta data.
One of the next steps is fine-tuning which is required for the parameter selection of the individual layers and we need to work on some thresholds regarding the utilization of the meta data. Nevertheless, the results are quite promising and we will continue to work on a self-contained prototype for movie and TV recommendations.