skus = set([sku for s in sessions for sku in s])
encoder = {y: x for x, y in enumerate(skus)}
decoder = [y for x, y in enumerate(skus)]

sku 是set, 编码器和解码的 index不是一个啊。。。这个地方有问题吧？？？
skus = sorted(set([sku for s in sessions for sku in s]))

是不是应该这样？？

Hi, I have a question about newtorkx graph and dgl graph.

The number of items is 33434. (0 ~ 33433) after encoding.
These items are entered into newtorkx's graph with some dropped items due to session length. so # of networkx's item is 14651.

that networkx's graph is transformed to dgl's graph. the items are relabeled from 0 ~ 14650. that dgl graph is 'train_g'.

dataloader get train_g's items like range(14651). so the dataloader's output is 0 ~ 14650.

but, in EGES model, item embedding is set by nn.Embedding(33434, 16) though dataloader's output is 0 ~ 14650.

is that right? but it works well.....

only 0 ~ 14650 index embeddings are used in embedding matrix, right?

for i in range(len(session) - 1):
edge = str(session[i]) + "," + str(session[i + 1])
try:
graph[edge] += 1
except KeyError:
graph[edge] = 1

构件图的时候使用的有向图吧？针对sku浏览行为使用无向图是不是更好些？？
毕竟看了A洗发水再看B洗发水 和 先看B洗发水再看A洗发水 对用户来说没有任何区别
是不是这样才对？？？？
for i in range(len(session) - 1):
# Ensure the smaller ID comes first for undirected edges
edge = tuple(sorted([session[i], session[i + 1]]))
try:
graph[edge] += 1
except KeyError:
graph[edge] = 1