This is an older note:

I think using graphs for recommendation systems is not a natural choice where temporal patterns and sequences play a role. It is not natural to represent time with nodes that can have multiple edges. Instead, RNNs or Markov chains might be more suitable. Recommendation is almost always sequential. Otherwise, it is about finding embeddings between elements. If so, graphs might not fit these problems. If graphs are not suitable, why do we use them?

The answer is simple: People don’t buy things in a strict sequence; they have options, multiple choices, and various factors that play a role in these choices. So, it is actually natural to model these interactions with graphs.