dataset_file.md

Dataset file

In order to use your own data, you have to create a dictionary with the following data format and compress it as a joblib dump file.

Standard GCN

"dense_adj" (required for GCN)

• a list of adjacency matrices.

"adj" [optional, alternative to "dense_adj" ] (required for GCN)

• Format: a list of a sparse adjacency matrix.
• A sparse matrix is represented as a tuple ('ind', 'value', 'shape'), where 'ind' expresses the indices of the matrix as a pair of row-col vectors (rows, cols), 'value' is a vector of the entries of the matrix, and 'shape' is a shape of the matrix, that is a pair of the number of rows and the number of cols.

"max_node_num" (required for GCN)

• Format: a scalar value of the maximum number of nodes in a graph.

"feature" (required for GCN with feature)

• Format: a list of M by D feature matrices (D is the number of features per node).

"label" (required for supervised training (of graph-centric GCN))

• Format: a list of E binary label matrices (E is the number of classes).

"node_num" [optional, node embedding mode]

• Format: a scalar value of the number of all nodes in all graph (= N)

"node" [optional, node embedding mode]

• Format: a list of a vector for indices of nodes in a graph. (0<= node index < N)

Multimodal

The following optoins are optional for multimodal mode (e.g. GCN and DNN)

"sequence"

• Format: a list of symbolic sequences as a integer matrix (the number of graphs x the maximum length of sequences)
• Each element is represented as an integer encoding a symbol (1<= element <=S).

"sequence_length"

• Format: a list of lengths of sequences. A length of this list should be the number of graphs.

"sequence_symbol_num"

• Format: a scalar value of the number of symbols in sequences (= S).

"sequence"

• Format: a list of symbolic sequences as a integer matrix (the number of graphs x the maximum length of sequences)
• Each element is represented as an integer encoding a symbol (1<= element <=S).

"profeat"/"dragon"/"ecfp"

• Format: a list of vectors as a floating matrix (the number of graphs x the dimension of features)
• "profeat", "dragon", and "ecfp" are processed as the same way.