Interaction

Interaction is the internal data structural that is loaded from DataLoader, and fed into the recommendation algorithms.

It is implemented as a new abstract data type based on python.dict. The keys correspond to features from input, which can be conveniently referenced with feature names when writing the recommendation algorithms; and the values correspond to tensors (implemented by torch.Tensor), which will be used for the update and computation in learning algorithms. Specially, the value entry for a specific key stores all the corresponding tensor data in a batch or mini-batch.

With such a data structure, our library provides a friendly interface to write the recommendation algorithms in a batch-based mode. For example, we can read all the user embeddings and item embeddings from an instantiated Interaction object inter simply based on the feature names:

user_vec = inter['UserID']
item_vec = inter['ItemID']

The contents of an Interaction are decided by the loaded fields. However, it should be noted that there can be some features generated by DataLoader, e.g. if one model has input_type = InputType.PAIRWISE, then each item feature has a corresponding negative item feature, whose keys are begin with arg NEG_PREFIX.

Besides, the value components are implemented based on torch.Tensor. We wrap many functions of PyTorch to develop a GRU-oriented data structure, which can support batch-based mechanism (e.g., copying a batch of data to GPU). In specific, we summarize the important functions as follows:

Function	Description
to(device)	transfer all tensors to torch.device
cpu	transfer all tensors to CPU
numpy	transfer all tensors to numpy.ndarray
repeat	repeats each tensor along the batch size dimension
repeat interleave	repeat elements of a tensor, similar to repeat interleave
update	update this object with another Interaction, similar to update