Why not consider the cosine similarity between elements and focus on developing better feature vectors? Possibly models trained with a focus on DOM semantics an

Why not consider the cosine similarity between elements and focus on developing better feature vectors? Possibly models trained with a focus on DOM semantics and graph structure?