How to input edge weight for PinSAGE evaluation step

connie · June 26, 2023, 7:24am

Hi, I’m trying to build the evaluation code based on below github.

dmlc/dgl/blob/7b1639f1aacb006fa65ef8cef09c49f5219bd5c1/examples/pytorch/pinsage/evaluation.py#L36

      
        
                self.user_to_item_etype = user_to_item_etype
                self.batch_size = batch_size
                self.timestamp = timestamp
            
            
def recommend(self, full_graph, K, h_user, h_item):
                """
                Return a (n_user, K) matrix of recommended items for each user
                """
                graph_slice = full_graph.edge_type_subgraph([self.user_to_item_etype])
                n_users = full_graph.num_nodes(self.user_ntype)
                latest_interactions = dgl.sampling.select_topk(
                    graph_slice, 1, self.timestamp, edge_dir="out"
                )
                user, latest_items = latest_interactions.all_edges(
                    form="uv", order="srcdst"
                )
                # each user should have at least one "latest" interaction
                assert torch.equal(user, torch.arange(n_users))
            
            
    recommended_batches = []
                user_batches = torch.arange(n_users).split(self.batch_size)

I’m trying to do it my own data instead of Movielens dataset.
Instead of data-dict[‘timestamp’] on Movielens data, how can I use the ‘timestamp’ of my graph data as the edge weight?

below screenshot is the graph data g which I created.

when I tried this code : g.edges[‘visited’].data[‘timestamp’]
I got the below error.

BarclayII · June 26, 2023, 10:00am

It seems that you are passing in a tensor rather than a string to dgl.sampling.select_topk. Instead, you need to call it like

dgl.sampling.select_topk(graph_slice, k, 'timestamp', ...)

where 'timestamp' is the name of your edge weight in the graph.

connie · July 3, 2023, 9:25am

thanks for your response. I resolved this issue. however, for the prediction, I got the below error… Could you advise?

BarclayII · July 3, 2023, 9:58am

It seems that your ground truth matrix is still a scipy COO matrix, which does not support indexing. You will need to convert it to a CSR matrix using .tocsr() method. You can see how I did it in evaluate_nn function in the same file.