Data

UsersInteractions

class recoder.data.UsersInteractions(users, interactions_matrix)

Holds the interactions of a set of users in an interactions sparse matrix

Parameters:

• users (np.array) – users being represented.
• interactions_matrix (scipy.sparse.csr_matrix) – user-item interactions matrix, where interactions_matrix[i] correspond to the interactions of users[i].

RecommendationDataset

class recoder.data.RecommendationDataset(interactions_matrix, target_interactions_matrix=None)

Represents a torch.utils.data.Dataset that iterates through the users interactions with items.

Indexing this dataset returns a UsersInteractions containing the interactions of the users in the index.

Parameters:

• interactions_matrix (scipy.sparse.csr_matrix) – the user-item interactions matrix.
• target_interactions_matrix (scipy.sparse.csr_matrix, optional) – the target user-item interactions matrix. Mainly used for evaluation, representing the items to recommend.

RecommendationDataLoader

class recoder.data.RecommendationDataLoader(dataset, batch_size, negative_sampling=False, num_sampling_users=0, num_workers=0, collate_fn=None)

A DataLoader similar to torch.utils.data.DataLoader that handles RecommendationDataset and generate batches with negative sampling.

By default, if no collate_fn is provided, the BatchCollator.collate() will be used, and iterating through this dataloader will return a Batch at each iteration.

Parameters:

• dataset (RecommendationDataset) – dataset from which to load the data
• batch_size (int) – number of samples per batch
• negative_sampling (bool, optional) – whether to apply mini-batch based negative sampling or not.
• num_sampling_users (int, optional) – number of users to consider for mini-batch based negative sampling. This is useful for increasing the number of negative samples while keeping the batch-size small. If 0, then num_sampling_users will be equal to batch_size.
• num_workers (int, optional) – how many subprocesses to use for data loading.
• collate_fn (callable, optional) – A function that transforms a UsersInteractions into a mini-batch.

Batch

class recoder.data.Batch(users, items, indices, values, size)

Represents a sparse batch of users and items interactions.

Parameters:

• users (torch.LongTensor) – users that are in the batch
• items (torch.LongTensor) – items that are in the batch
• indices (torch.LongTensor) – the indices of the interactions in the sparse matrix
• values (torch.LongTensor) – the values of the interactions
• size (torch.Size) – the size of the sparse interactions matrix

BatchCollator

class recoder.data.BatchCollator(batch_size, negative_sampling=False)

Collator of UsersInteractions. It collates the users interactions into multiple Batch based on batch_size.

Parameters:

• batch_size (int) – number of samples per batch
• negative_sampling (bool, optional) – whether to apply mini-batch based negative sampling or not.

collate(users_interactions)

Collates UsersInteractions into batches of size batch_size.

Parameters:users_interactions (UsersInteractions) – a UsersInteractions. Returns:list of batches. Return type:list[Batch]