Merge pull request #74 from downeykking/main

FEA: add DirectAU and fix some bugs

recbole_gnn/data/dataset.py

@@ -6,15 +6,33 @@
 from tqdm import tqdm
 from torch_geometric.utils import degree
 
+
 from recbole.data.dataset import SequentialDataset
 from recbole.data.dataset import Dataset as RecBoleDataset
 from recbole.utils import set_color, FeatureSource
 
+import recbole
+import pickle
+from recbole.utils import ensure_dir
+
 
 class GeneralGraphDataset(RecBoleDataset):
     def __init__(self, config):
         super().__init__(config)
 
+    if recbole.__version__ == "1.1.1":
+
+        def save(self):
+            """Saving this :class:`Dataset` object to :attr:`config['checkpoint_dir']`."""
+            save_dir = self.config["checkpoint_dir"]
+            ensure_dir(save_dir)
+            file = os.path.join(save_dir, f'{self.config["dataset"]}-{self.__class__.__name__}.pth')
+            self.logger.info(
+                set_color("Saving filtered dataset into ", "pink") + f"[{file}]"
+            )
+            with open(file, "wb") as f:
+                pickle.dump(self, f)
+
     def get_norm_adj_mat(self):
         r"""Get the normalized interaction matrix of users and items.
         Construct the square matrix from the training data and normalize it
@@ -101,6 +119,7 @@ def build(self):
             dataset.session_graph_construction()
         return datasets
 
+
 class MultiBehaviorDataset(SessionGraphDataset):
 
     def session_graph_construction(self):
@@ -113,7 +132,7 @@ def session_graph_construction(self):
             # To be compatible with existing datasets
             item_behavior_seq = torch.tensor([0] * len(item_seq))
             self.behavior_id_field = 'behavior_id'
-            self.field2id_token[self.behavior_id_field] = {0:'interaction'}
+            self.field2id_token[self.behavior_id_field] = {0: 'interaction'}
         else:
             item_behavior_seq = self.inter_feat[self.item_list_length_field]
 
@@ -152,6 +171,7 @@ def session_graph_construction(self):
             'alias_inputs': alias_inputs
         }
 
+
 class LESSRDataset(SessionGraphDataset):
     def session_graph_construction(self):
         self.logger.info('Constructing LESSR session graphs.')
@@ -199,14 +219,14 @@ def reverse_session(self):
         item_seq = self.inter_feat[self.item_id_list_field]
         item_seq_len = self.inter_feat[self.item_list_length_field]
         for i in tqdm(range(item_seq.shape[0])):
-            item_seq[i,:item_seq_len[i]] = item_seq[i,:item_seq_len[i]].flip(dims=[0])
+            item_seq[i, :item_seq_len[i]] = item_seq[i, :item_seq_len[i]].flip(dims=[0])
 
     def bidirectional_edge(self, edge_index):
         seq_len = edge_index.shape[1]
         ed = edge_index.T
         ed2 = edge_index.T.flip(dims=[1])
         idc = ed.unsqueeze(1).expand(-1, seq_len, 2) == ed2.unsqueeze(0).expand(seq_len, -1, 2)
-        return torch.logical_and(idc[:,:,0], idc[:,:,1]).any(dim=-1)
+        return torch.logical_and(idc[:, :, 0], idc[:, :, 1]).any(dim=-1)
 
     def session_graph_construction(self):
         self.logger.info('Constructing session graphs.')
@@ -276,9 +296,10 @@ class SocialDataset(GeneralGraphDataset):
         net_feat (pandas.DataFrame): Internal data structure stores the users' social network relations.
             It's loaded from file ``.net``.
     """
+
     def __init__(self, config):
         super().__init__(config)
-    
+
     def _get_field_from_config(self):
         super()._get_field_from_config()
 
@@ -410,4 +431,4 @@ def net_matrix(self, form='coo', value_field=None):
         Returns:
             scipy.sparse: Sparse matrix in form ``coo`` or ``csr``.
         """
-        return self._create_sparse_matrix(self.net_feat, self.net_src_field, self.net_tgt_field, form, value_field)
+        return self._create_sparse_matrix(self.net_feat, self.net_src_field, self.net_tgt_field, form, value_field)

recbole_gnn/model/general_recommender/__init__.py

@@ -6,3 +6,4 @@
 from recbole_gnn.model.general_recommender.lightgcl import LightGCL
 from recbole_gnn.model.general_recommender.simgcl import SimGCL
 from recbole_gnn.model.general_recommender.xsimgcl import XSimGCL
+from recbole_gnn.model.general_recommender.directau import DirectAU

recbole_gnn/model/general_recommender/directau.py

@@ -0,0 +1,120 @@
+# r"""
+# DiretAU
+# ################################################
+# Reference:
+#     Chenyang Wang et al. "Towards Representation Alignment and Uniformity in Collaborative Filtering." in KDD 2022.
+
+# Reference code:
+#     https://github.com/THUwangcy/DirectAU
+# """
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from recbole.model.init import xavier_normal_initialization
+from recbole.utils import InputType
+from recbole.model.general_recommender import BPR
+from recbole_gnn.model.general_recommender import LightGCN
+
+from recbole_gnn.model.abstract_recommender import GeneralGraphRecommender
+
+
+class DirectAU(GeneralGraphRecommender):
+    input_type = InputType.PAIRWISE
+
+    def __init__(self, config, dataset):
+        super(DirectAU, self).__init__(config, dataset)
+
+        # load parameters info
+        self.embedding_size = config['embedding_size']
+        self.gamma = config['gamma']
+        self.encoder_name = config['encoder']
+
+        # define encoder
+        if self.encoder_name == 'MF':
+            self.encoder = MFEncoder(config, dataset)
+        elif self.encoder_name == 'LightGCN':
+            self.encoder = LGCNEncoder(config, dataset)
+        else:
+            raise ValueError('Non-implemented Encoder.')
+
+        # storage variables for full sort evaluation acceleration
+        self.restore_user_e = None
+        self.restore_item_e = None
+
+        # parameters initialization
+        self.apply(xavier_normal_initialization)
+
+    def forward(self, user, item):
+        user_e, item_e = self.encoder(user, item)
+        return F.normalize(user_e, dim=-1), F.normalize(item_e, dim=-1)
+
+    @staticmethod
+    def alignment(x, y, alpha=2):
+        return (x - y).norm(p=2, dim=1).pow(alpha).mean()
+
+    @staticmethod
+    def uniformity(x, t=2):
+        return torch.pdist(x, p=2).pow(2).mul(-t).exp().mean().log()
+
+    def calculate_loss(self, interaction):
+        if self.restore_user_e is not None or self.restore_item_e is not None:
+            self.restore_user_e, self.restore_item_e = None, None
+
+        user = interaction[self.USER_ID]
+        item = interaction[self.ITEM_ID]
+
+        user_e, item_e = self.forward(user, item)
+        align = self.alignment(user_e, item_e)
+        uniform = self.gamma * (self.uniformity(user_e) + self.uniformity(item_e)) / 2
+
+        return align, uniform
+
+    def predict(self, interaction):
+        user = interaction[self.USER_ID]
+        item = interaction[self.ITEM_ID]
+        user_e = self.user_embedding(user)
+        item_e = self.item_embedding(item)
+        return torch.mul(user_e, item_e).sum(dim=1)
+
+    def full_sort_predict(self, interaction):
+        user = interaction[self.USER_ID]
+        if self.encoder_name == 'LightGCN':
+            if self.restore_user_e is None or self.restore_item_e is None:
+                self.restore_user_e, self.restore_item_e = self.encoder.get_all_embeddings()
+            user_e = self.restore_user_e[user]
+            all_item_e = self.restore_item_e
+        else:
+            user_e = self.encoder.user_embedding(user)
+            all_item_e = self.encoder.item_embedding.weight
+        score = torch.matmul(user_e, all_item_e.transpose(0, 1))
+        return score.view(-1)
+
+
+class MFEncoder(BPR):
+    def __init__(self, config, dataset):
+        super(MFEncoder, self).__init__(config, dataset)
+
+    def forward(self, user_id, item_id):
+        return super().forward(user_id, item_id)
+
+    def get_all_embeddings(self):
+        user_embeddings = self.user_embedding.weight
+        item_embeddings = self.item_embedding.weight
+        return user_embeddings, item_embeddings
+
+
+class LGCNEncoder(LightGCN):
+    def __init__(self, config, dataset):
+        super(LGCNEncoder, self).__init__(config, dataset)
+
+    def forward(self, user_id, item_id):
+        user_all_embeddings, item_all_embeddings = self.get_all_embeddings()
+        u_embed = user_all_embeddings[user_id]
+        i_embed = item_all_embeddings[item_id]
+        return u_embed, i_embed
+
+    def get_all_embeddings(self):
+        return super().forward()

recbole_gnn/properties/model/DirectAU.yaml

@@ -0,0 +1,7 @@
+embedding_size: 64
+encoder: "MF"   # "MF" or "lightGCN"
+gamma: 0.5
+weight_decay: 1e-6
+train_batch_size: 256
+
+# n_layers: 3 # needed for LightGCN

recbole_gnn/quick_start.py

@@ -80,6 +80,7 @@ def objective_function(config_dict=None, config_file_list=None, saved=True):
     test_result = trainer.evaluate(test_data, load_best_model=saved)
 
     return {
+        'model': config['model'],
         'best_valid_score': best_valid_score,
         'valid_score_bigger': config['valid_metric_bigger'],
         'best_valid_result': best_valid_result,

tests/test_model.py

@@ -60,7 +60,7 @@ def test_simgcl(self):
             'model': 'SimGCL'
         }
         quick_test(config_dict)
-        
+
     def test_xsimgcl(self):
         config_dict = {
             'model': 'XSimGCL'
@@ -73,6 +73,12 @@ def test_lightgcl(self):
         }
         quick_test(config_dict)
 
+    def test_directau(self):
+        config_dict = {
+            'model': 'DirectAU'
+        }
+        quick_test(config_dict)
+
 
 class TestSequentialRecommender(unittest.TestCase):
     def test_gru4rec(self):