retrieval.py
Is the main file for getting the top k songs based on one input song. Execute the file in the terminal with $python retrieval.py ,and enter the artist name, song title and number of songs to be retrieved. Output is a list of songs in the format ['artist', 'song_title']. The type of measure to use and which features to consider for recommendation can be chosen in the file. features = 'best' computes the ranking according to the method described in the report. Main retrieval file. Files for best method
creating_genre_dict.py
Creates a pickled dictionary of the genres in the format {ID: [Genres]} for the whole id_genres_mmsr.tsv file. Should be executed before executing evaluation.py. File is stored under 'genre_dictionary.pkl'
create_top_k_dict.py
Creates a pickled dictionary according to a given measure (default is an average of the cosine similarity over tf-idf, BERT and word2vec, other options are a jaccard similarity over tf-idf(not recommended due to computation time) or an average of the inner product over tf-idf, BERT and word2vec). Calculates the top 100 songs for each song in the files and stores their IDs in the format {SongID: [Top1ID, Top2ID,...]} . Retrieved IDs are stored after every 1000th song. File is stored under [measure]+'_retrieved_ids.pkl'. Basically same functionality as retrieval.py but it is adapted to more efficiently input all the songs as retrieval.py is more to be used using song title and artist name.
A precomputed version for all songs using the different modalities can be downloaded here: Top 100 retrievals. You can find this file in the computing_dicts branch.
evaluation.py
Evaluates the file created in create_top_k_dict.py and prints the average precision, MMR and nDCG at 10 and 100 songs.
PCA.ipynb
Jupyter notebook to create the dimensionality reduction of the different datafiles using PCA.
baseline.ipynb
Jupyter notebook to create the baseline dictionary, where for each songs the a song with the genre rock is recommended for each k.
genre_analysis.ipynb
Jupyter notebook of the data analysis.
kendall_tau_correlation.ipynb
Jupyter notebook for computing the kendall tau correlation and creating the plots.
precision_recall_plots.ipynb
Jupyter notebook to efficiently compute the precision and recall for different k and creating the precision-recall plot.
popularity.ipynb
Jupyter notebook to compute the percent delta mean for k10 and k100. The results for percent delta mean(q) can be found here: percent delta mean(q) results
hubness.ipynb
Jupyter note book to compute the hubness metric for k=10 and k=100.
late-fusion.ipynb
Jupyter notebook to compute the simple or weighted borda count on precomputed rankings.
Group members:
Rahil Mujadidi, k11904249
Birgit Reiter, k12209548
Martin Seidl, k11908861
Sebastian Wolff, k12007396
A̶g̶n̶e̶s̶ ̶H̶i̶n̶t̶e̶r̶p̶l̶a̶t̶t̶n̶e̶r̶,̶ ̶k̶0̶1̶6̶3̶5̶1̶8̶3̶