Big impact from item features

MLK · December 22, 2021, 12:17pm

Hello,

I’m using Pinsage implementation from dgl. I’ve got item features (image embeddings) that I’m using as a FloatTensor (node features) in a similar way like that:

github.com

dmlc/dgl/blob/master/examples/pytorch/pinsage/process_nowplaying_rs.py#L51

    
      
          float_cols = []
          for col in tracks.columns:
              if col == 'track_id':
                  continue
              elif col == 'artist_id':
                  g.nodes['track'].data[col] = torch.LongTensor(tracks[col].cat.codes.values)
              elif tracks.dtypes[col] == 'float64':
                  float_cols.append(col)
              else:
                  g.nodes['track'].data[col] = torch.LongTensor(tracks[col].values)
          g.nodes['track'].data['song_features'] = torch.FloatTensor(linear_normalize(tracks[float_cols].values))
          g.edges['listened'].data['created_at'] = torch.LongTensor(events['created_at'].values)
          g.edges['listened-by'].data['created_at'] = torch.LongTensor(events['created_at'].values)
          
          
n_edges = g.number_of_edges('listened')
          train_indices, val_indices, test_indices = train_test_split_by_time(events, 'created_at', 'user_id')
          train_g = build_train_graph(g, train_indices, 'user', 'track', 'listened', 'listened-by')
          assert train_g.out_degrees(etype='listened').min() > 0
          val_matrix, test_matrix = build_val_test_matrix(
              g, val_indices, test_indices, 'user', 'track', 'listened')

I saw that they have a big impact in that line, especially when they weren’t normalized:

github.com

dmlc/dgl/blob/master/examples/pytorch/pinsage/layers.py#L109

    
      
          
          
            module = self.inputs[feature]
                      if isinstance(module, (BagOfWords, BagOfWordsPretrained)):
                          # Textual feature; find the length and pass it to the textual module.
                          length = ndata[feature + '__len']
                          result = module(data, length)
                      else:
                          result = module(data)
                      projections.append(result)
          
          
        return torch.stack(projections, 1).sum(1)
          
          
class WeightedSAGEConv(nn.Module):
              def __init__(self, input_dims, hidden_dims, output_dims, act=F.relu):
                  super().__init__()
          
          
        self.act = act
                  self.Q = nn.Linear(input_dims, hidden_dims)
                  self.W = nn.Linear(input_dims + hidden_dims, output_dims)
                  self.reset_parameters()
                  self.dropout = nn.Dropout(0.5)

I think that item features have a bigger impact than IDs embeddings. Should I change something to reduce impact from image embeddings and base more on interactions (user - item)? I think that in a recommendation system it could be a better idea.

Best regards
K

BarclayII · December 24, 2021, 5:20am

Whether to reduce the impact mostly depends on your dataset and scenario; sometimes image features can be quite discriminative and sometimes it may be not. That being said, how did you measure the “impact”? Did you compare the model performance between adding and not adding the image features? If so, you should handle features the way you get better performance.

system · January 23, 2022, 5:20am

This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.