请教下为啥无监督类的训练用两个网络分别encoder这么设计？

当前看walk系列的算法包括hne 都是 src encoder一个网络， pos、negs encoder一个网络。这么做的好处是啥，为啥不用一个网络？ 可以用一个网络吗？

另外如果这么设计我可不可以用，pos&negs的网络参数作为emb的结果，话说这俩网络原理上啥区别？

Start by tracing the walk and HNE implementations under src, comparing the encoder used for the source node with the encoders for pos and neg samples. Check how each output is consumed as the final embedding and document the roles, trade-offs, parameter-sharing options, and which output should represent an embedding. Done means these questions are answered with references to the relevant entry points.

Written by the indexing model from the issue text.