In real world settings it’s common to have features describing each item as well as ratings or other counts expressing users historical interactions with items. As we expect that users might like items with similar features to those that they have liked in the past, it should be useful for a recommender to take item features into account. One way of doing this is to extend the matrix factorization approach, which represents each user and item with a low-dimensional vector, by learning to represent each feature in the same low-dimensional space. mrec includes an implementation of a joint model of this kind which optimizes the WARP ranking loss [1]. This model learns an embedding matrix which maps the item features into the low-dimensional space. To predict the rating or preference score for an unseen item, it computes the dot product of the user factor and item factor in the usual way for a matrix factorization recommender, but then adds the dot product of the user factor and the low-dimensional mapping of its feature vector.
As an example we’ll look again at the small movie ratings dataset that we previously worked with in Getting started with mrec, but this time we’ll add some features based on movie plot descriptions from IMDb. To create the features, first download the plot.list.gz file from one of the official IMDb ftp sites. This contains plot summaries for most of the movies in the MovieLens datasets. Once you’ve unzipped this file you can use the extract_movie_features script in the bin directory of the mrec source tree to create features and save them to file:
$ cd mrec
$ ./bin/extract_movie_features plot.list ml-100k/u.item 100k.features.npz
Note
The extract_movie_features script isn’t installed automatically with mrec so you’ll need to grab the source code if you don’t already have it.
The resulting features are simply tf-idf counts of the words found in the plot summaries for each movie. You can load them like this:
>>> from mrec import load_sparse_matrix
>>> features = load_sparse_matrix('npz','100k.features.npz')
and inspect the top few word counts for the first few items:
>>> for i in xrange(3):
... for tfidf,word in sorted(zip(features[i].data,features[i].indices),reverse=True)[:3]:
... print '{0}\t{1}\t{2:.3f}'.format(i,word,tfidf)
...
0 500 0.440
0 549 0.340
0 4 0.242
1 311 0.412
1 564 0.335
1 549 0.243
2 117 0.430
2 286 0.427
2 670 0.220
Now we can train a recommender in the usual way, specifying the features with the item_features and item_feature_format options:
$ mrec_train -n4 --input_format tsv --train u.data.train.0 --outdir models --model warp --item_features 100k.features.npz --item_feature_format npz
Once this has finished (it will take a few minutes even on a single split of this small dataset) you can use the recommender to make and evaluate predictions:
$ mrec_predict --input_format tsv --test_input_format tsv --train u.data.train.0 --modeldir models --outdir recs --item_features 100k.features.npz --item_feature_format npz
After a few seconds you’ll get the results as usual:
WARP2MF(d=80,gamma=0.01,C=100.0)
mrr 0.6008 +/- 0.0000
prec@5 0.3650 +/- 0.0000
prec@10 0.3221 +/- 0.0000
prec@15 0.2915 +/- 0.0000
prec@20 0.2699 +/- 0.0000
| [1] | Weston, J., Bengio, S., & Usunier, N. (2010). Large scale image annotation: learning to rank with joint word-image embeddings. Machine learning, 81(1), 21-35. |