nextrec 0.4.1__py3-none-any.whl → 0.4.2__py3-none-any.whl

nextrec/models/ranking/autoint.py

@@ -3,8 +3,8 @@ Date: create on 09/11/2025
 Checkpoint: edit on 24/11/2025
 Author: Yang Zhou,zyaztec@gmail.com
 Reference:
-[1] Song W, Shi C, Xiao Z, et al. Autoint: Automatic feature interaction learning via 
-self-attentive neural networks[C]//Proceedings of the 28th ACM international conference 
+[1] Song W, Shi C, Xiao Z, et al. Autoint: Automatic feature interaction learning via
+self-attentive neural networks[C]//Proceedings of the 28th ACM international conference
 on information and knowledge management. 2019: 1161-1170.
 (https://arxiv.org/abs/1810.11921)
 
@@ -70,29 +70,31 @@ class AutoInt(BaseModel):
     @property
     def default_task(self):
         return "binary"
-    
-    def __init__(self,
-                 dense_features: list[DenseFeature],
-                 sparse_features: list[SparseFeature],
-                 sequence_features: list[SequenceFeature],
-                 att_layer_num: int = 3,
-                 att_embedding_dim: int = 8,
-                 att_head_num: int = 2,
-                 att_dropout: float = 0.0,
-                 att_use_residual: bool = True,
-                 target: list[str] | None = None,
-                 task: str | list[str] | None = None,
-                 optimizer: str = "adam",
-                 optimizer_params: dict | None = None,
-                 loss: str | nn.Module | None = "bce",
-                 loss_params: dict | list[dict] | None = None,
-                 device: str = 'cpu',
-                 embedding_l1_reg=1e-6,
-                 dense_l1_reg=1e-5,
-                 embedding_l2_reg=1e-5,
-                 dense_l2_reg=1e-4,
-                 **kwargs):
-        
+
+    def __init__(
+        self,
+        dense_features: list[DenseFeature],
+        sparse_features: list[SparseFeature],
+        sequence_features: list[SequenceFeature],
+        att_layer_num: int = 3,
+        att_embedding_dim: int = 8,
+        att_head_num: int = 2,
+        att_dropout: float = 0.0,
+        att_use_residual: bool = True,
+        target: list[str] | None = None,
+        task: str | list[str] | None = None,
+        optimizer: str = "adam",
+        optimizer_params: dict | None = None,
+        loss: str | nn.Module | None = "bce",
+        loss_params: dict | list[dict] | None = None,
+        device: str = "cpu",
+        embedding_l1_reg=1e-6,
+        dense_l1_reg=1e-5,
+        embedding_l2_reg=1e-5,
+        dense_l2_reg=1e-4,
+        **kwargs,
+    ):
+
         super(AutoInt, self).__init__(
             dense_features=dense_features,
             sparse_features=sparse_features,
@@ -104,7 +106,7 @@ class AutoInt(BaseModel):
             dense_l1_reg=dense_l1_reg,
             embedding_l2_reg=embedding_l2_reg,
             dense_l2_reg=dense_l2_reg,
-            **kwargs
+            **kwargs,
         )
 
         if target is None:
@@ -113,52 +115,59 @@ class AutoInt(BaseModel):
             optimizer_params = {}
         if loss is None:
             loss = "bce"
-            
+
         self.att_layer_num = att_layer_num
         self.att_embedding_dim = att_embedding_dim
-        
+
         # Use sparse and sequence features for interaction
         # **INFO**: this is different from the original paper, we also include dense features
         # if you want to follow the paper strictly, set dense_features=[]
         # or modify the code accordingly
-        self.interaction_features = dense_features + sparse_features + sequence_features 
-        
+        self.interaction_features = dense_features + sparse_features + sequence_features
+
         # All features for embedding
         self.all_features = dense_features + sparse_features + sequence_features
 
         # Embedding layer
         self.embedding = EmbeddingLayer(features=self.all_features)
-        
+
         # Project embeddings to attention embedding dimension
         num_fields = len(self.interaction_features)
-        
+
         # If embeddings have different dimensions, project them to att_embedding_dim
-        self.need_projection = not all(f.embedding_dim == att_embedding_dim for f in self.interaction_features)
+        self.need_projection = not all(
+            f.embedding_dim == att_embedding_dim for f in self.interaction_features
+        )
         self.projection_layers = None
         if self.need_projection:
-            self.projection_layers = nn.ModuleList([
-                nn.Linear(f.embedding_dim, att_embedding_dim, bias=False) 
-                for f in self.interaction_features
-            ])
-        
+            self.projection_layers = nn.ModuleList(
+                [
+                    nn.Linear(f.embedding_dim, att_embedding_dim, bias=False)
+                    for f in self.interaction_features
+                ]
+            )
+
         # Multi-head self-attention layers
-        self.attention_layers = nn.ModuleList([
-            MultiHeadSelfAttention(
-                embedding_dim=att_embedding_dim,
-                num_heads=att_head_num,
-                dropout=att_dropout,
-                use_residual=att_use_residual
-            ) for _ in range(att_layer_num)
-        ])
-        
+        self.attention_layers = nn.ModuleList(
+            [
+                MultiHeadSelfAttention(
+                    embedding_dim=att_embedding_dim,
+                    num_heads=att_head_num,
+                    dropout=att_dropout,
+                    use_residual=att_use_residual,
+                )
+                for _ in range(att_layer_num)
+            ]
+        )
+
         # Final prediction layer
         self.fc = nn.Linear(num_fields * att_embedding_dim, 1)
         self.prediction_layer = PredictionLayer(task_type=self.default_task)
 
         # Register regularization weights
         self.register_regularization_weights(
-            embedding_attr='embedding',
-            include_modules=['projection_layers', 'attention_layers', 'fc']
+            embedding_attr="embedding",
+            include_modules=["projection_layers", "attention_layers", "fc"],
         )
 
         self.compile(
@@ -172,21 +181,29 @@ class AutoInt(BaseModel):
         # Get embeddings field-by-field so mixed dimensions can be projected safely
         field_embeddings = []
         if len(self.interaction_features) == 0:
-            raise ValueError("AutoInt requires at least one sparse or sequence feature for interactions.")
+            raise ValueError(
+                "AutoInt requires at least one sparse or sequence feature for interactions."
+            )
         for idx, feature in enumerate(self.interaction_features):
             feature_emb = self.embedding(x=x, features=[feature], squeeze_dim=False)
             feature_emb = feature_emb.squeeze(1)  # [B, embedding_dim]
             if self.need_projection and self.projection_layers is not None:
                 feature_emb = self.projection_layers[idx](feature_emb)
-            field_embeddings.append(feature_emb.unsqueeze(1))  # [B, 1, att_embedding_dim or original_dim]
+            field_embeddings.append(
+                feature_emb.unsqueeze(1)
+            )  # [B, 1, att_embedding_dim or original_dim]
         embeddings = torch.cat(field_embeddings, dim=1)
-        
+
         # Apply multi-head self-attention layers
         attention_output = embeddings
         for att_layer in self.attention_layers:
-            attention_output = att_layer(attention_output)  # [B, num_fields, att_embedding_dim]
-        
+            attention_output = att_layer(
+                attention_output
+            )  # [B, num_fields, att_embedding_dim]
+
         # Flatten and predict
-        attention_output_flat = attention_output.flatten(start_dim=1)  # [B, num_fields * att_embedding_dim]
+        attention_output_flat = attention_output.flatten(
+            start_dim=1
+        )  # [B, num_fields * att_embedding_dim]
         y = self.fc(attention_output_flat)  # [B, 1]
         return self.prediction_layer(y)

nextrec/models/ranking/dcn.py

@@ -15,21 +15,27 @@ from nextrec.basic.model import BaseModel
 from nextrec.basic.layers import EmbeddingLayer, MLP, PredictionLayer
 from nextrec.basic.features import DenseFeature, SparseFeature, SequenceFeature
 
+
 class CrossNetwork(nn.Module):
     """Stacked Cross Layers from DCN (Wang et al., 2017)."""
 
     def __init__(self, input_dim, num_layers):
         super().__init__()
         self.num_layers = num_layers
-        self.w = torch.nn.ModuleList([torch.nn.Linear(input_dim, 1, bias=False) for _ in range(num_layers)])
-        self.b = torch.nn.ParameterList([torch.nn.Parameter(torch.zeros((input_dim,))) for _ in range(num_layers)])
+        self.w = torch.nn.ModuleList(
+            [torch.nn.Linear(input_dim, 1, bias=False) for _ in range(num_layers)]
+        )
+        self.b = torch.nn.ParameterList(
+            [torch.nn.Parameter(torch.zeros((input_dim,))) for _ in range(num_layers)]
+        )
 
     def forward(self, x):
         x0 = x
         for i in range(self.num_layers):
             xw = self.w[i](x)
             x = x0 * xw + self.b[i] + x
-        return x # [batch_size, input_dim]
+        return x  # [batch_size, input_dim]
+
 
 class DCN(BaseModel):
     @property
@@ -40,25 +46,27 @@ class DCN(BaseModel):
     def default_task(self):
         return "binary"
 
-    def __init__(self,
-                 dense_features: list[DenseFeature],
-                 sparse_features: list[SparseFeature],
-                 sequence_features: list[SequenceFeature],
-                 cross_num: int = 3,
-                 mlp_params: dict | None = None,
-                 target: list[str] = [],
-                 task: str | list[str] | None = None,
-                 optimizer: str = "adam",
-                 optimizer_params: dict = {},
-                 loss: str | nn.Module | None = "bce",
-                 loss_params: dict | list[dict] | None = None,
-                 device: str = 'cpu',
-                 embedding_l1_reg=1e-6,
-                 dense_l1_reg=1e-5,
-                 embedding_l2_reg=1e-5,
-                 dense_l2_reg=1e-4,
-                 **kwargs):
-        
+    def __init__(
+        self,
+        dense_features: list[DenseFeature],
+        sparse_features: list[SparseFeature],
+        sequence_features: list[SequenceFeature],
+        cross_num: int = 3,
+        mlp_params: dict | None = None,
+        target: list[str] = [],
+        task: str | list[str] | None = None,
+        optimizer: str = "adam",
+        optimizer_params: dict = {},
+        loss: str | nn.Module | None = "bce",
+        loss_params: dict | list[dict] | None = None,
+        device: str = "cpu",
+        embedding_l1_reg=1e-6,
+        dense_l1_reg=1e-5,
+        embedding_l2_reg=1e-5,
+        dense_l2_reg=1e-4,
+        **kwargs,
+    ):
+
         super(DCN, self).__init__(
             dense_features=dense_features,
             sparse_features=sparse_features,
@@ -70,13 +78,13 @@ class DCN(BaseModel):
             dense_l1_reg=dense_l1_reg,
             embedding_l2_reg=embedding_l2_reg,
             dense_l2_reg=dense_l2_reg,
-            **kwargs
+            **kwargs,
         )
 
         self.loss = loss
         if self.loss is None:
             self.loss = "bce"
-            
+
         # All features
         self.all_features = dense_features + sparse_features + sequence_features
 
@@ -84,20 +92,30 @@ class DCN(BaseModel):
         self.embedding = EmbeddingLayer(features=self.all_features)
 
         # Calculate input dimension
-        emb_dim_total = sum([f.embedding_dim for f in self.all_features if not isinstance(f, DenseFeature)])
-        dense_input_dim = sum([getattr(f, "embedding_dim", 1) or 1 for f in dense_features])
+        emb_dim_total = sum(
+            [
+                f.embedding_dim
+                for f in self.all_features
+                if not isinstance(f, DenseFeature)
+            ]
+        )
+        dense_input_dim = sum(
+            [getattr(f, "embedding_dim", 1) or 1 for f in dense_features]
+        )
         input_dim = emb_dim_total + dense_input_dim
-        
+
         # Cross Network
         self.cross_network = CrossNetwork(input_dim=input_dim, num_layers=cross_num)
-        
+
         # Deep Network (optional)
         if mlp_params is not None:
             self.use_dnn = True
             self.mlp = MLP(input_dim=input_dim, **mlp_params)
             deep_dim = self.mlp.output_dim
             # Final layer combines cross and deep
-            self.final_layer = nn.Linear(input_dim + deep_dim, 1)  # + deep_dim for MLP output
+            self.final_layer = nn.Linear(
+                input_dim + deep_dim, 1
+            )  # + deep_dim for MLP output
         else:
             self.use_dnn = False
             # Final layer only uses cross network output
@@ -107,8 +125,8 @@ class DCN(BaseModel):
 
         # Register regularization weights
         self.register_regularization_weights(
-            embedding_attr='embedding',
-            include_modules=['cross_network', 'mlp', 'final_layer']
+            embedding_attr="embedding",
+            include_modules=["cross_network", "mlp", "final_layer"],
         )
 
         self.compile(
@@ -121,18 +139,20 @@ class DCN(BaseModel):
     def forward(self, x):
         # Get all embeddings and flatten
         input_flat = self.embedding(x=x, features=self.all_features, squeeze_dim=True)
-        
+
         # Cross Network
         cross_output = self.cross_network(input_flat)  # [B, input_dim]
-        
+
         if self.use_dnn:
             # Deep Network
             deep_output = self.mlp(input_flat)  # [B, 1]
             # Concatenate cross and deep
-            combined = torch.cat([cross_output, deep_output], dim=-1)  # [B, input_dim + 1]
+            combined = torch.cat(
+                [cross_output, deep_output], dim=-1
+            )  # [B, input_dim + 1]
         else:
             combined = cross_output
-        
+
         # Final prediction
         y = self.final_layer(combined)
         return self.prediction_layer(y)

nextrec/models/ranking/dcn_v2.py

@@ -5,25 +5,30 @@ Date: create on 09/11/2025
 import torch
 import torch.nn as nn
 
-from nextrec.basic.model import BaseModel
-from nextrec.basic.layers import EmbeddingLayer, MLP, PredictionLayer
-from nextrec.basic.features import DenseFeature, SparseFeature, SequenceFeature
 
 class CrossNetV2(nn.Module):
     """Vector-wise cross network proposed in DCN V2 (Wang et al., 2021)."""
+
     def __init__(self, input_dim, num_layers):
         super().__init__()
         self.num_layers = num_layers
-        self.w = torch.nn.ModuleList([torch.nn.Linear(input_dim, input_dim, bias=False) for _ in range(num_layers)])
-        self.b = torch.nn.ParameterList([torch.nn.Parameter(torch.zeros((input_dim,))) for _ in range(num_layers)])
-
+        self.w = torch.nn.ModuleList(
+            [
+                torch.nn.Linear(input_dim, input_dim, bias=False)
+                for _ in range(num_layers)
+            ]
+        )
+        self.b = torch.nn.ParameterList(
+            [torch.nn.Parameter(torch.zeros((input_dim,))) for _ in range(num_layers)]
+        )
 
     def forward(self, x):
         x0 = x
         for i in range(self.num_layers):
-            x =x0*self.w[i](x) + self.b[i] + x
+            x = x0 * self.w[i](x) + self.b[i] + x
         return x
-    
+
+
 class CrossNetMix(nn.Module):
     """Mixture of low-rank cross experts from DCN V2 (Wang et al., 2021)."""
 
@@ -33,18 +38,46 @@ class CrossNetMix(nn.Module):
         self.num_experts = num_experts
 
         # U: (input_dim, low_rank)
-        self.u_list = torch.nn.ParameterList([nn.Parameter(nn.init.xavier_normal_(
-            torch.empty(num_experts, input_dim, low_rank))) for i in range(self.num_layers)])
+        self.u_list = torch.nn.ParameterList(
+            [
+                nn.Parameter(
+                    nn.init.xavier_normal_(
+                        torch.empty(num_experts, input_dim, low_rank)
+                    )
+                )
+                for i in range(self.num_layers)
+            ]
+        )
         # V: (input_dim, low_rank)
-        self.v_list = torch.nn.ParameterList([nn.Parameter(nn.init.xavier_normal_(
-            torch.empty(num_experts, input_dim, low_rank))) for i in range(self.num_layers)])
+        self.v_list = torch.nn.ParameterList(
+            [
+                nn.Parameter(
+                    nn.init.xavier_normal_(
+                        torch.empty(num_experts, input_dim, low_rank)
+                    )
+                )
+                for i in range(self.num_layers)
+            ]
+        )
         # C: (low_rank, low_rank)
-        self.c_list = torch.nn.ParameterList([nn.Parameter(nn.init.xavier_normal_(
-            torch.empty(num_experts, low_rank, low_rank))) for i in range(self.num_layers)])
-        self.gating = nn.ModuleList([nn.Linear(input_dim, 1, bias=False) for i in range(self.num_experts)])
-
-        self.bias = torch.nn.ParameterList([nn.Parameter(nn.init.zeros_(
-            torch.empty(input_dim, 1))) for i in range(self.num_layers)])
+        self.c_list = torch.nn.ParameterList(
+            [
+                nn.Parameter(
+                    nn.init.xavier_normal_(torch.empty(num_experts, low_rank, low_rank))
+                )
+                for i in range(self.num_layers)
+            ]
+        )
+        self.gating = nn.ModuleList(
+            [nn.Linear(input_dim, 1, bias=False) for i in range(self.num_experts)]
+        )
+
+        self.bias = torch.nn.ParameterList(
+            [
+                nn.Parameter(nn.init.zeros_(torch.empty(input_dim, 1)))
+                for i in range(self.num_layers)
+            ]
+        )
 
     def forward(self, x):
         x_0 = x.unsqueeze(2)  # (bs, in_features, 1)
@@ -59,7 +92,9 @@ class CrossNetMix(nn.Module):
 
                 # (2) E(x_l)
                 # project the input x_l to $\mathbb{R}^{r}$
-                v_x = torch.matmul(self.v_list[i][expert_id].t(), x_l)  # (bs, low_rank, 1)
+                v_x = torch.matmul(
+                    self.v_list[i][expert_id].t(), x_l
+                )  # (bs, low_rank, 1)
 
                 # nonlinear activation in low rank space
                 v_x = torch.tanh(v_x)
@@ -67,7 +102,9 @@ class CrossNetMix(nn.Module):
                 v_x = torch.tanh(v_x)
 
                 # project back to $\mathbb{R}^{d}$
-                uv_x = torch.matmul(self.u_list[i][expert_id], v_x)  # (bs, in_features, 1)
+                uv_x = torch.matmul(
+                    self.u_list[i][expert_id], v_x
+                )  # (bs, in_features, 1)
 
                 dot_ = uv_x + self.bias[i]
                 dot_ = x_0 * dot_  # Hadamard-product
@@ -75,10 +112,14 @@ class CrossNetMix(nn.Module):
                 output_of_experts.append(dot_.squeeze(2))
 
             # (3) mixture of low-rank experts
-            output_of_experts = torch.stack(output_of_experts, 2)  # (bs, in_features, num_experts)
-            gating_score_experts = torch.stack(gating_score_experts, 1)  # (bs, num_experts, 1)
+            output_of_experts = torch.stack(
+                output_of_experts, 2
+            )  # (bs, in_features, num_experts)
+            gating_score_experts = torch.stack(
+                gating_score_experts, 1
+            )  # (bs, num_experts, 1)
             moe_out = torch.matmul(output_of_experts, gating_score_experts.softmax(1))
             x_l = moe_out + x_l  # (bs, in_features, 1)
 
         x_l = x_l.squeeze()  # (bs, in_features)
-        return x_l
+        return x_l

nextrec/models/ranking/deepfm.py

@@ -43,13 +43,13 @@ embedding，无需手工构造交叉特征即可端到端训练，常用于 CTR/
 - CTR/CVR 任务的常用强基线
 """
 
-import torch
 import torch.nn as nn
 
 from nextrec.basic.model import BaseModel
 from nextrec.basic.layers import FM, LR, EmbeddingLayer, MLP, PredictionLayer
 from nextrec.basic.features import DenseFeature, SparseFeature, SequenceFeature
 
+
 class DeepFM(BaseModel):
     @property
     def model_name(self):
@@ -59,23 +59,26 @@ class DeepFM(BaseModel):
     def default_task(self):
         return "binary"
 
-    def __init__(self,
-                 dense_features: list[DenseFeature]|list = [],
-                 sparse_features: list[SparseFeature]|list = [],
-                 sequence_features: list[SequenceFeature]|list = [],
-                 mlp_params: dict = {},
-                 target: list[str]|str = [],
-                 task: str | list[str] | None = None,
-                 optimizer: str = "adam",
-                 optimizer_params: dict = {},
-                 loss: str | nn.Module | None = "bce",
-                 loss_params: dict | list[dict] | None = None,
-                 device: str = 'cpu',
-                 embedding_l1_reg=1e-6,
-                 dense_l1_reg=1e-5,
-                 embedding_l2_reg=1e-5,
-                 dense_l2_reg=1e-4, **kwargs):
-        
+    def __init__(
+        self,
+        dense_features: list[DenseFeature] | list = [],
+        sparse_features: list[SparseFeature] | list = [],
+        sequence_features: list[SequenceFeature] | list = [],
+        mlp_params: dict = {},
+        target: list[str] | str = [],
+        task: str | list[str] | None = None,
+        optimizer: str = "adam",
+        optimizer_params: dict = {},
+        loss: str | nn.Module | None = "bce",
+        loss_params: dict | list[dict] | None = None,
+        device: str = "cpu",
+        embedding_l1_reg=1e-6,
+        dense_l1_reg=1e-5,
+        embedding_l2_reg=1e-5,
+        dense_l2_reg=1e-4,
+        **kwargs,
+    ):
+
         super(DeepFM, self).__init__(
             dense_features=dense_features,
             sparse_features=sparse_features,
@@ -87,13 +90,13 @@ class DeepFM(BaseModel):
             dense_l1_reg=dense_l1_reg,
             embedding_l2_reg=embedding_l2_reg,
             dense_l2_reg=dense_l2_reg,
-            **kwargs
+            **kwargs,
         )
 
         self.loss = loss
         if self.loss is None:
             self.loss = "bce"
-            
+
         self.fm_features = sparse_features + sequence_features
         self.deep_features = dense_features + sparse_features + sequence_features
         self.embedding = EmbeddingLayer(features=self.deep_features)
@@ -107,8 +110,15 @@ class DeepFM(BaseModel):
         self.prediction_layer = PredictionLayer(task_type=self.default_task)
 
         # Register regularization weights
-        self.register_regularization_weights(embedding_attr='embedding', include_modules=['linear', 'mlp'])
-        self.compile(optimizer=optimizer, optimizer_params=optimizer_params, loss=loss, loss_params=loss_params)
+        self.register_regularization_weights(
+            embedding_attr="embedding", include_modules=["linear", "mlp"]
+        )
+        self.compile(
+            optimizer=optimizer,
+            optimizer_params=optimizer_params,
+            loss=loss,
+            loss_params=loss_params,
+        )
 
     def forward(self, x):
         input_deep = self.embedding(x=x, features=self.deep_features, squeeze_dim=True)