Model & Train

Pure and Feat Models

LibRecommender is a hybrid recommender system, which means you can choose whether to use features other than user behaviors or not. For models only use user behaviors, we classify them as pure models. This category includes UserCF, ItemCF, SVD, SVD++, ALS, NCF, BPR, RNN4Rec, Item2Vec, Caser, WaveNet, DeepWalk, NGCF, LightGCN.

Then for models that can use other features (e.g., age, sex, name etc.), we call them feat models. This category includes WideDeep, FM, DeepFM, YouTubeRetrieval, YouTubeRanking, AutoInt, DIN, GraphSage, PinSage, TwoTower.

The main difference on usage between these two kinds of models are:

1. pure models should use DatasetPure to process data, and feat models should use DatasetFeat.

2. When using feat models, four parameters should be provided, i.e. [sparse_col, dense_col, user_col, item_col], as otherwise the model will have no idea how to deal with all kinds of features.

The fit() method is the sole method for training a model in LibRecommender. You can find some typical usages in these examples:

See also

• pure_rating_example.py

• pure_ranking_example.py

• feat_rating_example.py

• feat_ranking_example.py

In fact, there exists two other kinds of model categories in LibRecommender, and we call them sequence and graph models. You can find them in the algorithm list.

Sequence models leverage information of user behavior sequence, whereas Graph models leverage information from graph. As you can see, these models overlap with pure and feat models. But no need to worry, the APIs remain the same, and you can just refer to the examples above.

Multiprocess data loading

For most models, users can leverage torch.utils.data.DataLoader to enable multiprocess data loading and speed up training. Users can specify the number of workers to be used during data loading using the num_workers parameter. Please refer to the corresponding PyTorch documentation for more information.

>>> model.fit(train_data, neg_sampling=False, num_workers=2)

Loss

LibRecommender provides some options on loss type for ranking task. The default loss type for most models is cross entropy loss. Since version 0.10.0, focal loss was added into the library. First introduced in Lin et al., 2018, focal loss down-weights well-classified examples and focuses on hard examples to get better training performance, and here is the implementation.

In order to choose which loss to use, simply set the loss_type parameter:

>>> model = Caser(task="ranking", loss_type="cross_entropy", ...)
>>> model = Caser(task="ranking", loss_type="focal", ...)

The table below lists the losses and negative samplers that can be used for ranking task in each algorithm:

Algorithm	Loss	Sampler
UserCF, ItemCF, ALS, Item2Vec, DeepWalk	/	/
BPR	bpr	random, unconsumed, popular
YouTubeRetrieval	sampled_softmax, nce	uniform, other
SVD, SVD++, NCF, Wide&Deep, FM, DeepFM, YouTubeRanking, AutoInt, DIN, Caser, WaveNet	cross_entropy, focal	random, unconsumed, popular
RNN4Rec	cross_entropy, focal, bpr	random, unconsumed, popular
NGCF, LightGCN	cross_entropy, focal, bpr, max_margin	random, unconsumed, popular
GraphSage, GraphSageDGL, PinSage, PinSageDGL	cross_entropy, focal, bpr, max_margin	random, unconsumed, popular, out-batch
TwoTower	cross_entropy, max_margin, softmax	random, unconsumed, popular

Caution

bpr and max_margin belong to pairwise loss, so they must be used with negative sampling, which means your data should only contains positive samples when using these losses.