Metapath2Vec
metapath2vec.RdRuns metapath2vec over a heterogeneous graph. Random walks are constrained
to follow a metapath schema over node types, e.g.
c("gene", "pathway", "gene"), and a skip-gram model is trained on them.
With metapath_plus = TRUE, negative samples are drawn from the context
node's own type (metapath2vec++).
Walks start only on nodes of the metapath's first type. Nodes no walk ever
reaches keep their random initialisation, so by default they are removed
and listed in the unvisited_nodes attribute. The function warns when more
than 10% of walks are dropped or walks reach less than half the requested
length on average: the schema does not fit the graph. Details are in the
walk_stats attribute. A metapath over a single type warns as well; that
is DeepWalk on a subgraph, so use node2vec().
Usage
metapath2vec(
graph_dt,
node_dt,
metapath,
embd_dim = 8L,
metapath2vec_params = params_metapath2vec(),
metapath_plus = FALSE,
filter_unvisited = TRUE,
directed = FALSE,
seed = 42L,
.verbose = TRUE
)Arguments
- graph_dt
data.table. The edge table. Needs to have the columns
"from"and"to", and can optionally have a"weight"column.- node_dt
data.table. The node table with the columns
"id"and"type". Every edge endpoint needs to be inid.- metapath
Character vector. The metapath over node types, closing on its starting type, e.g.
c("gene", "pathway", "gene").- embd_dim
Integer. Size of the embedding dimensions to create. Defaults to
8L.- metapath2vec_params
Named list. The training parameters, see
params_metapath2vec().- metapath_plus
Boolean. Use per-type negative sampling (metapath2vec++). Defaults to
FALSE.- filter_unvisited
Boolean. Remove the rows of nodes that no walk visited. Defaults to
TRUE.- directed
Boolean. Indicates if this is a directed or undirected network. Defaults to
FALSE.- seed
Integer. Seed for reproducibility.
- .verbose
Boolean. Controls verbosity of the function.
Value
A numeric matrix of n_nodes x embd_dim with the node ids as
rownames. Rows are ordered by node type. Carries three attributes:
node_type - Named factor with the type of each row.
unvisited_nodes - Character vector with the ids of the nodes no walk visited. Removed from the matrix if
filter_unvisited = TRUE.walk_stats - List with
start_nodes,attempted,truncated,dropped,mean_lengthandwalk_length.