The transformation of data into other forms of representation has played a major role in building the core programmatic systems of the Top-n Music Genre Classification Neural Network as well as the content recommendation system that utilizes the trained neural network for data preprocessing and recommendation generation. The way my team’s neural network operates is by taking as an input an audio clip and returning a list of a selected number of genres sorted in order of descending confidence percentage. In total, our model is trained on ten different genres meaning that each prediction created with our model will produce a list of ten percentage values for each genre.
A dilemma I faced when beginning the process of conceptualizing a content recommendation system was that content can be suggested based on the similarity of the content according to some metric or the similarity of the users that engage with the content. The latter solution, while very compelling, seems to require some form of preexisting organic user ecosystem to become functional. I decided to try to solve for the former type of recommendation system after realizing that I can use the resulting list, or vector, produced by our model as a form of unique profile for each audio clip in our database. By comparing the prediction vector produced for a novel audio clip by our model to each preprocessed audio clip vector in our recommendation database utilizing an element-wise difference of vectors I was able to extend the core technology developed for our project for the purpose of drawing mathematical connections between different audio data.
Working on this project stretch goal has allowed me to reflect on the importance of selecting the best fitting data abstraction for the task at hand especially if the chosen data abstraction does not contain a large amount of the potential information or features that can be extracted from the subject being analyzed.
Leave a Reply