Description

You will do this assignment in groups of three. You can seek partners via Piazza. Course staff can also help you find partners. Submit one PDF writeup per team via the Canvas site.
For this practical, you are tasked with predicting people’s tastes in music. Specifically, you will be predicting the number of times (a non-negative integer) different users listened to tracks of different artists over a span of time. You will have some basic selfreported demographic information about many, but not all, of the users, such as sex, age, and location. You will also have the name of the artist and their MusicBrainz ID, if available. There are about 233,000 users and 2,000 artists. The training set has over 4.1M user/artist pairs and the test set is of a similar size. Your objective is to predict how many times a user will listen to a new artist.
Data Files
There are five data files of interest, which can be downloaded from https://inclass. kaggle.com/c/cs181-s16-practical-3-predicting-music-tastes/data:
• profiles.csv – This file contains information about the users. There is a header row and then four columns in basic CSV format with comma delimiters and doublequote escaping where appropriate. The first few rows are:
user,sex,age,country fa40b43298ba3f8aa52e8e8863faf2e2171e0b5d,f,25,Sweden 5909125332c108365a26ccf0ee62636eee08215c,m,29,Iceland d1867cbda35e0d48e9a8390d9f5e079c9d99ea96,m,30,United States 63268cce0d68127729890c1691f62d5be5abd87c,m,21,Germany
02871cd952d607ba69b64e2e107773012c708113,m,24,Netherlands 0938eb3d1b449b480c4e2431c457f6ead7063a34,m,22,United States e4c6b36e65db3d48474dd538fe74d2dbb5a2e79e,f,,United States b97479f9a563a5c43b423a976f51fd509e1ec5ba,f,,Poland 3bb020df0ff376dfdded4d5e63e2d35a50b3c535,m,,United States f3fb86c0f024f640cae3fb479f3a27e0dd499891,,16,Ukraine …
• artists.csv – This file contains information about the 2,000 artists that have been listened to in these data. There is a header row and then five columns. The first several rows are:
artist,name 03098741-08b3-4dd7-b3f6-1b0bfa2c879c,Liars
7a2e6b55-f149-4e74-be6a-30a1b1a387bb,The Desert Sessions
7002bf88-1269-4965-a772-4ba1e7a91eaa,Glenn Gould dbf7c761-e332-467b-b4d9-aafe06bbcf8f,G. Love & Special Sauce a3cb23fc-acd3-4ce0-8f36-1e5aa6a18432,U2 8b0f05ce-354e-4121-9e0b-8b4732ea844f,Juanes
8363f94f-fd86-41b8-a56b-26eacb34f499,Summoning 2e41ae9c-afd2-4f20-8f1e-17281ce9b472,Gwen Stefani c17f08f4-2542-46fb-97f3-3202d60c225a,Fear Factory 4bd95eea-b9f6-4d70-a36c-cfea77431553,Alice in Chains f467181e-d5e0-4285-b47e-e853dcc89ee7,Ratatat …
The first column is the MusicBrainz ID. The second is the name, if available.
• train.csv – This file contains the training data, which are 4.1M artist/user pairs with numbers of plays. It is a standard CSV file with comma delimiters and a header row. The first column is the user identifier, followed by the artist identifier, and then the number of plays. Example rows:
user,artist,plays eb1c … af03,5a8e … 9c94,554 44ce … bb5d,a3a9 … 84df,81 da9c … 08e3,eeb1 … 8e43,708 8fa4 … 7d81,a141 … eabc,265 b85f … b2cf,a3cb … 8432,220 feed … 08f7,1cc5 … f506,2113 cbb8 … 324b,9c9f … d090,127 5641 … 3e9b,832a … 1c24,305 …
• test.csv – This file is a CSV file which contains users and artists, but without the plays column. Your objective is to predict these values and create a prediction file, which is described below. Each user/artist pair has a distinct Id; you’ll need to match this to your predictions. There are 4,154,805 pairs to predict. About half of these are used to compute the visible leaderboard. The other half are used to compute the true results. This separation is to prevent overfitting to the leaderboard, and is standard for these kinds of prediction contests. The first couple of rows of the test file are:
Id,user,artist
1,306e … 22d2,4ac4e32b-bd18-402e-adad-ae00e72f8d85
2,9450 … 27ac,1f574ab1-a46d-4586-9331-f0ded23e0411
3,8019 … 53cc,3eb72791-6322-466b-87d3-24d74901eb2d
4,e3ed … 5d82,61604b45-8a91-4e33-a1b6-45d7b1fec4e5
5,a73f … 44aa,5dfdca28-9ddc-4853-933c-8bc97d87beec
6,55f1 … c0af,ef58d4c9-0d40-42ba-bfab-9186c1483edd
7,7ad7 … bbee,a3cb23fc-acd3-4ce0-8f36-1e5aa6a18432 …
• global_median.csv – This is an example of how you submit predictions. It is a standard comma-delimited CSV file with two columns. The Id column corresponds to entries in the test.csv file above, i.e., specific user/artist pairs. The plays column is where you specify your best guess. Although the true number of plays are integers, you can produce floating point numbers in your predictions. An example is below:
Id,plays
1,118.0
2,118.0
3,118.0
4,118.0
5,118.0
6,118.0
7,118.0
8,118.0
9,118.0
10,118.0
11,118.0
12,118.0
13,118.0
14,118.0 …
Evaluation
After you upload your predictions to Kaggle (which you can do at most four times per day), they will be compared to the held-out true number of plays. The score is computed via mean absolute error (lower is better). If there are N test data, where your prediction is xˆn and the truth is xn, then the MAE is
1 N MAE = ∑ |xˆn − xn|
N n=1
Sample Code
Two Python files are available, which can be downloaded from https://inclass. kaggle.com/c/cs181-s16-practical-3-predicting-music-tastes/data. The files global_median.py and user_median.py implement rudimentary predictions based on simple ideas.
Solution Ideas
As in the previous practicals, you have a lot of flexibility in what you might do. You could focus on feature engineering, i.e., coming up with fancy inputs for your method, or you could focus on fancy techniques that use the features. You could also really drive the practical with unsupervised learning and use clustering techniques or matrix factorization. Here are some ideas to get you started:
• Goldberg, Ken, Theresa Roeder, Dhruv Gupta, and Chris Perkins. “Eigentaste: A constant time collaborative filtering algorithm.” Information Retrieval 4, no. 2 (2001): 133-151.
• Mnih, Andriy, and Ruslan Salakhutdinov. “Probabilistic matrix factorization”. In Advances in Neural Information Processing Systems, pp. 1257-1264. 2007.
• Koren, Yehuda, Robert Bell, and Chris Volinsky. “Matrix factorization techniques for recommender systems.” Computer 42, no. 8 (2009): 30-37.