Application to collaborative filtering

Collaborative filtering

This package contains all the necessary functions to quickly train a model for a collaborative filtering task. Let's start by importing all we'll need.

from fastai.collab import * 


Collaborative filtering is when you're tasked to predict how much a user is going to like a certain item. The fastai library contains a CollabFilteringDataset class that will help you create datasets suitable for training, and a function get_colab_learner to build a simple model directly from a ratings table. Let's first see how we can get started before delving into the documentation.

For this example, we'll use a small subset of the MovieLens dataset to predict the rating a user would give a particular movie (from 0 to 5). The dataset comes in the form of a csv file where each line is a rating of a movie by a given person.

path = untar_data(URLs.ML_SAMPLE)
ratings = pd.read_csv(path/'ratings.csv')
userId movieId rating timestamp
0 73 1097 4.0 1255504951
1 561 924 3.5 1172695223
2 157 260 3.5 1291598691
3 358 1210 5.0 957481884
4 130 316 2.0 1138999234

We'll first turn the userId and movieId columns in category codes, so that we can replace them with their codes when it's time to feed them to an Embedding layer. This step would be even more important if our csv had names of users, or names of items in it. To do it, we simply have to call a CollabDataBunch factory method.

data = CollabDataBunch.from_df(ratings)

Now that this step is done, we can directly create a Learner object:

learn = collab_learner(data, n_factors=50, y_range=(0.,5.))

And then immediately begin training

learn.fit_one_cycle(5, 5e-3, wd=0.1)
Total time: 00:09

epoch train_loss valid_loss
1 2.427430 1.999472
2 1.116335 0.663345
3 0.736155 0.636640
4 0.612827 0.626773
5 0.565003 0.626336

class CollabDataBunch[source]

CollabDataBunch(train_dl:DataLoader, valid_dl:DataLoader, fix_dl:DataLoader=None, test_dl:Optional[DataLoader]=None, device:device=None, dl_tfms:Optional[Collection[Callable]]=None, path:PathOrStr='.', collate_fn:Callable='data_collate', no_check:bool=False) :: DataBunch

Base DataBunch for collaborative filtering.

The init function shouldn't be called directly (as it's the one of a basic DataBunch), instead, you'll want to use the following factory method.


from_df(ratings:DataFrame, pct_val:float=0.2, user_name:Optional[str]=None, item_name:Optional[str]=None, rating_name:Optional[str]=None, test:DataFrame=None, seed:int=None, path:PathOrStr='.', bs:int=64, val_bs:int=None, num_workers:int=4, dl_tfms:Optional[Collection[Callable]]=None, device:device=None, collate_fn:Callable='data_collate', no_check:bool=False) → CollabDataBunch

Create a DataBunch suitable for collaborative filtering from ratings.

Take a ratings dataframe and splits it randomly for train and test following pct_val (unless it's None). user_name, item_name and rating_name give the names of the corresponding columns (defaults to the first, the second and the third column). Optionally a test dataframe can be passed an a seed for the separation between training and validation set. The kwargs will be passed to DataBunch.create.

Model and Learner

class CollabLearner[source]

CollabLearner(data:DataBunch, model:Module, opt_func:Callable='Adam', loss_func:Callable=None, metrics:Collection[Callable]=None, true_wd:bool=True, bn_wd:bool=True, wd:Floats=0.01, train_bn:bool=True, path:str=None, model_dir:str='models', callback_fns:Collection[Callable]=None, callbacks:Collection[Callback]=<factory>, layer_groups:ModuleList=None) :: Learner

Learner suitable for collaborative filtering.

This is a subclass of Learner that just introduces helper functions to analyze results, the initialization is the same as a regular Learner.


bias(arr:Collection[T_co], is_item:bool=True)

Bias for item or user (based on is_item) for all in arr. (Set model to cpu and no grad.)


get_idx(arr:Collection[T_co], is_item:bool=True)

Fetch item or user (based on is_item) for all in arr. (Set model to cpu and no grad.)


weight(arr:Collection[T_co], is_item:bool=True)

Bias for item or user (based on is_item) for all in arr. (Set model to cpu and no grad.)

class EmbeddingDotBias[source]

EmbeddingDotBias(n_factors:int, n_users:int, n_items:int, y_range:Point=None) :: Module

Base dot model for collaborative filtering.

Creates a simple model with Embedding weights and biases for n_users and n_items, with n_factors latent factors. Takes the dot product of the embeddings and adds the bias, then if y_range is specified, feed the result to a sigmoid rescaled to go from y_range[0] to y_range[1].

class EmbeddingNN[source]

EmbeddingNN(emb_szs:ListSizes, layers:Collection[int]=None, ps:Collection[float]=None, emb_drop:float=0.0, y_range:OptRange=None, use_bn:bool=True, bn_final:bool=False) :: TabularModel

Subclass TabularModel to create a NN suitable for collaborative filtering.

emb_szs will overwrite the default and kwargs are passed to TabularModel.


collab_learner(data, n_factors:int=None, use_nn:bool=False, emb_szs:Dict[str, int]=None, layers:Collection[int]=None, ps:Collection[float]=None, emb_drop:float=0.0, y_range:OptRange=None, use_bn:bool=True, bn_final:bool=False, **learn_kwargs) → Learner

Create a Learner for collaborative filtering on data.

More specifically, binds data with a model that is either an EmbeddingDotBias with n_factors if use_nn=False or a EmbeddingNN with emb_szs otherwise. In both cases the numbers of users and items will be inferred from the data, y_range can be specified in the kwargs and you can pass metrics or wd to the Learner constructor.

class CollabLine[source]

CollabLine(cats, conts, classes, names) :: TabularLine

Subclass of TabularLine for collaborative filtering.

class CollabList[source]

CollabList(items:Iterator[T_co], cat_names:OptStrList=None, cont_names:OptStrList=None, procs=None, **kwargs) → TabularList :: TabularList

Subclass of TabularList for collaborative filtering.