This is a technique for finding continuous vector-space embeddings of heterogeneous networks. The basic idea is much the same as node2vec and word2vec—it still relies on the skip-gram model. However, this new appraoch requires the user to specify a "meta-path", a path of node/edge types, which biases the random walk towards certain categories. For example, a heterogeneous academic network may contain Authors (A), Publications (P), and Venues (V); these might be linked in terms of publication (A, P), co-authorship (A-A), and venue publication (P, V). We could then define a meta-path, such as APA, that will bias the random walker to start with an Author, transition to a Paper, and then transition to another Author. Similarly, other metapaths might include APVPA, Author-Paper-Venue-Paper-Author. Using these metapath-based random walks, we can embed heterogeneous networks. There is also metapath++, a procedure presented by the author that follows the same sampling/random walking procedure, but defines the output layer differently in the neural network. I still need help understanding this approach. The parameters are much the same as in other skip-gram based neural network procedures. However, metapath2vec requires that we define a metapath scheme upfront, which will require some sort of domain knolwedge/understnading of how things are related. For example, in the scholarly network example, APA signifies co-author relationships , whereas APVPA signified relationships between authors who publish in the same venue. There will need to be experimentation to get this right.