FROSTT file formats

We provide several file formats for tensors.

Sparse tensors (.tns)

Sparse tensors come in a text format which is human-readable and designed to be easy to generate and work with. Each line of the file represents a non-zero. An order-M tensor has M+1 entries per line (separated by whitespace). The first M entries specify the indices (or, coordinates) of the non-zero, and the last entry provides the non-zero value. Indices start from one.

Example

This is a fourth-order tensor with five non-zeros:

# This is a comment
1 1 1 1 1.0
1 1 2 1 2.0
2 2 2 1 3.0
2 3 1 2 4.0
3 1 1 2 5.0

Mode mappings (.map)

We provide maps of the tensor indices back to the original data whenever possible. This information is provided using a .map file. This file format is a text file with as many lines as the dimension of the tensor. Line i provides the ID of the ith index in the tensor.

Please note that we are not always able to provide mappings to the original data. This may be due to anonymization or because they simply were not shared with us.

Example

Suppose a tensor of product ratings has the first mode representing 2000 users, and the second mode representing 100 products. The file mode-1-users.map would have 2000 lines, with each line containing the username or ID of a customer. Similarly, the file mode-2-products.map would have 100 lines, each containing the name or ID of a product.