Music-Recommendation-System In this project, we analyzed sequential patterns of songs userwise. We took One Hot, Word2Vec and Bert word embeddings and used as an input representation of words. Finally, we made Horizontal CNN and Vertical CNN and concatenated them to get our final results.
- Developed a music recommendation system using next basket prediction technique.
- The system uses machine learning and Convolutional Neural Networks (CNN) techniques to suggest music to users based on their preferences and past behaviour.
- Includes CNN applied over basket and target with the help of different Embeddings.
- Achieved remarkable accuracy of 74.68% on 20k samples of data.
- In contrast to standard predictive models, recommender systems need training on data in a sequential manner, often called sessions.
- I implemented model by dividing data into unique users. Next for each user, I made many sessions for the user which shows many sequences of music that were of interest of the user.
- Then, I applied Ensembled CNN model in which I processed data in a Horizontal CNN model and a Vertical CNN model and then concatenated them to form a single but powerful learning model.
- Used adam as optimizer and categorical_crossentropy as loss for training model.
- Trained on One Hot Embedding, Word2Vec Embedding and BERT Embeddings giving results as (42.91% vs 63.27% vs 74.68%).
- Scoring of products user-wise and in the form of sequences known as baskets.
- Another edge, using Ensembled CNN model with pre-trained Embeddings for better understanding User's behaviour and preferences.
- Numpy
- Pandas
- Convolutional Neural Network (CNN)
- Tensorflow
- BERT and Word2Vec Embeddings
- Matplotlib
- Python
- Jupyter Notebooks
- Google Colab
- Github
I built Recommendation System based on a real-world dataset: 30Music. As entire dataset is very large (1.8 million samples of data), I used only a part of it for training model.
Data Source Link: https://recsys.deib.polimi.it/datasets/
Model Accuracy
