nextrec 0.4.1__py3-none-any.whl → 0.4.3__py3-none-any.whl

nextrec/models/ranking/dcn_v2.py

@@ -1,5 +1,46 @@
 """
 Date: create on 09/11/2025
+Checkpoint: edit on 09/12/2025
+Author: Yang Zhou, zyaztec@gmail.com
+Reference:
+[1] R. Wang et al. DCN V2: Improved Deep & Cross Network and Practical Lessons for
+Web-scale Learning to Rank Systems. KDD 2021.
+(https://arxiv.org/abs/2008.13535)
+
+DCN v2 enhances the original Deep & Cross Network by replacing the scalar cross
+weights with vector-wise (matrix) parameters and a Mixture-of-Low-Rank-Experts
+variant. The matrix cross (CrossNetV2) improves expressiveness with manageable
+parameter growth, while CrossNetMix decomposes the matrix into low-rank factors and
+gates across experts for stronger modeling at a similar cost. As in DCN, the cross
+tower explicitly builds polynomial feature interactions and can be paired with a
+deep MLP tower; their outputs are concatenated before a final linear head.
+
+Workflow:
+  (1) Embed sparse/sequence fields and concatenate with dense inputs
+  (2) Cross tower: choose matrix CrossNetV2 or low-rank CrossNetMix for explicit crosses
+  (3) Optional deep tower: MLP over the same flattened embeddings
+  (4) Fuse cross and deep outputs, then predict via a linear + prediction layer
+
+Key Advantages:
+- Vector-wise/matrix cross weights capture richer interactions than DCN v1
+- Low-rank MoE cross (CrossNetMix) boosts capacity without quadratic parameters
+- Compatible with a deep tower for additional nonlinear modeling
+
+DCN v2 在原始 DCN 基础上，将标量交叉权重升级为向量/矩阵参数，并引入低秩专家混合
+的 CrossNetMix。矩阵交叉（CrossNetV2）在参数可控的前提下提升表达力，CrossNetMix
+通过低秩分解和 gating 进一步增强建模能力且保持参数效率。和 DCN 一样，交叉塔显式
+构造多项式交互，可并行或串联一个 MLP 深塔，最终拼接/输出到线性头做预测。
+
+流程：
+  (1) 对稀疏/序列特征做 embedding，并与稠密特征拼接
+  (2) 交叉塔：可选矩阵 CrossNetV2 或低秩混合 CrossNetMix 显式构造交互
+  (3) 可选深塔：MLP 处理同一展平后的输入或交叉输出
+  (4) 融合交叉与深塔输出，经线性层和预测层得到最终得分
+
+主要优点：
+  - 矩阵交叉相较 DCN v1 捕获更丰富的交互
+  - 低秩专家混合在相近参数量下带来更强建模能力
+  - 兼容并行/串行深塔，灵活扩展非线性表示
 """
 
 import torch
@@ -9,76 +50,255 @@ 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):
+
+    def __init__(self, input_dim: int, num_layers: int):
         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):
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
         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)."""
 
-    def __init__(self, input_dim, num_layers=2, low_rank=32, num_experts=4):
-        super(CrossNetMix, self).__init__()
+    def __init__(
+        self,
+        input_dim: int,
+        num_layers: int = 2,
+        low_rank: int = 32,
+        num_experts: int = 4,
+    ):
+        super().__init__()
         self.num_layers = num_layers
         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)])
-        # 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)])
-        # 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)])
-
-    def forward(self, x):
+        self.u_list = nn.ParameterList(
+            [
+                nn.Parameter(
+                    nn.init.xavier_normal_(
+                        torch.empty(num_experts, input_dim, low_rank)
+                    )
+                )
+                for _ in range(num_layers)
+            ]
+        )
+        self.v_list = nn.ParameterList(
+            [
+                nn.Parameter(
+                    nn.init.xavier_normal_(
+                        torch.empty(num_experts, input_dim, low_rank)
+                    )
+                )
+                for _ in range(num_layers)
+            ]
+        )
+        self.c_list = nn.ParameterList(
+            [
+                nn.Parameter(
+                    nn.init.xavier_normal_(torch.empty(num_experts, low_rank, low_rank))
+                )
+                for _ in range(num_layers)
+            ]
+        )
+
+        self.gating = nn.ModuleList(
+            [nn.Linear(input_dim, 1, bias=False) for _ in range(num_experts)]
+        )
+
+        self.bias = nn.ParameterList(
+            [nn.Parameter(torch.zeros(input_dim, 1)) for _ in range(num_layers)]
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        # x: (bs, in_features)
         x_0 = x.unsqueeze(2)  # (bs, in_features, 1)
         x_l = x_0
+
         for i in range(self.num_layers):
             output_of_experts = []
             gating_score_experts = []
+
+            gating_input = x_l.squeeze(2)  # (bs, in_features)
+
             for expert_id in range(self.num_experts):
-                # (1) G(x_l)
-                # compute the gating score by x_l
-                gating_score_experts.append(self.gating[expert_id](x_l.squeeze(2)))
+                # Gating
+                gating_score_experts.append(
+                    self.gating[expert_id](gating_input)
+                )  # (bs, 1)
 
-                # (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)
+                # Low-rank cross: U C V^T x_l
+                V = self.v_list[i][expert_id]  # (in_features, low_rank)
+                C = self.c_list[i][expert_id]  # (low_rank, low_rank)
+                U = self.u_list[i][expert_id]  # (in_features, low_rank)
 
-                # nonlinear activation in low rank space
-                v_x = torch.tanh(v_x)
-                v_x = torch.matmul(self.c_list[i][expert_id], v_x)
+                # (bs, 1, low_rank)
+                v_x = x_l.transpose(1, 2).matmul(V)  # x_l^T V
+                v_x = v_x.matmul(C)  # · C
                 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)
+                # (bs, in_features, 1)
+                uv_x = U.matmul(v_x.transpose(1, 2))
+
+                # x_0 ⊙ (uv_x + b)
+                dot_ = x_0 * (uv_x + self.bias[i])  # (bs, in_features, 1)
+
+                output_of_experts.append(dot_.squeeze(2))  # (bs, in_features)
+
+            # (3) Mixture of experts
+            output_of_experts = torch.stack(
+                output_of_experts, dim=2
+            )  # (bs, in_features, num_experts)
+            gating_score_experts = torch.stack(
+                gating_score_experts, dim=1
+            )  # (bs, num_experts, 1)
+            gating_score_experts = gating_score_experts.softmax(dim=1)
+
+            moe_out = torch.matmul(
+                output_of_experts, gating_score_experts
+            )  # (bs, in_features, 1)
+            x_l = moe_out + x_l  # residual
+
+        return x_l.squeeze(-1)  # (bs, in_features)
+
+
+class DCNv2(BaseModel):
+    @property
+    def model_name(self) -> str:
+        return "DCNv2"
+
+    @property
+    def default_task(self):
+        return "binary"
+
+    def __init__(
+        self,
+        dense_features: list[DenseFeature] | None = None,
+        sparse_features: list[SparseFeature] | None = None,
+        sequence_features: list[SequenceFeature] | None = None,
+        cross_num: int = 3,
+        cross_type: str = "matrix",
+        architecture: str = "parallel",
+        low_rank: int = 32,
+        num_experts: int = 4,
+        mlp_params: dict | None = None,
+        target: list[str] | 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,
+    ):
+        dense_features = dense_features or []
+        sparse_features = sparse_features or []
+        sequence_features = sequence_features or []
+        optimizer_params = optimizer_params or {}
+        if loss is None:
+            loss = "bce"
+
+        super(DCNv2, self).__init__(
+            dense_features=dense_features,
+            sparse_features=sparse_features,
+            sequence_features=sequence_features,
+            target=target,
+            task=task or self.default_task,
+            device=device,
+            embedding_l1_reg=embedding_l1_reg,
+            dense_l1_reg=dense_l1_reg,
+            embedding_l2_reg=embedding_l2_reg,
+            dense_l2_reg=dense_l2_reg,
+            **kwargs,
+        )
+
+        self.all_features = dense_features + sparse_features + sequence_features
+        self.embedding = EmbeddingLayer(features=self.all_features)
+        input_dim = self.embedding.input_dim
+
+        architecture = architecture.lower()
+        if architecture not in {"parallel", "stacked"}:
+            raise ValueError("architecture must be 'parallel' or 'stacked'.")
+        self.architecture = architecture
+
+        cross_type = cross_type.lower()
+        if cross_type == "matrix":
+            self.cross_network = CrossNetV2(input_dim=input_dim, num_layers=cross_num)
+        elif cross_type in {"mix", "low_rank"}:
+            self.cross_network = CrossNetMix(
+                input_dim=input_dim,
+                num_layers=cross_num,
+                low_rank=low_rank,
+                num_experts=num_experts,
+            )
+        else:
+            raise ValueError("Unsupported cross_type for DCNv2. Use 'matrix' or 'mix'.")
+
+        if mlp_params is not None:
+            self.use_dnn = True
+            dnn_params = dict(mlp_params)
+            dnn_params.setdefault("output_layer", False)
+            self.mlp = MLP(input_dim=input_dim, **dnn_params)
+            deep_dim = self.mlp.output_dim
+            final_input_dim = (
+                input_dim + deep_dim if architecture == "parallel" else deep_dim
+            )
+        else:
+            if architecture == "stacked":
+                raise ValueError(
+                    "Stacked architecture requires mlp_params (deep tower)."
+                )
+            self.use_dnn = False
+            self.mlp = None
+            final_input_dim = input_dim
+
+        self.final_layer = nn.Linear(final_input_dim, 1)
+        self.prediction_layer = PredictionLayer(task_type=self.default_task)
+
+        self.register_regularization_weights(
+            embedding_attr="embedding",
+            include_modules=["cross_network", "mlp", "final_layer"],
+        )
 
-                dot_ = uv_x + self.bias[i]
-                dot_ = x_0 * dot_  # Hadamard-product
+        self.compile(
+            optimizer=optimizer,
+            optimizer_params=optimizer_params,
+            loss=loss,
+            loss_params=loss_params,
+        )
 
-                output_of_experts.append(dot_.squeeze(2))
+    def forward(self, x) -> torch.Tensor:
+        input_flat = self.embedding(x=x, features=self.all_features, squeeze_dim=True)
+        cross_out = self.cross_network(input_flat)
 
-            # (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)
-            moe_out = torch.matmul(output_of_experts, gating_score_experts.softmax(1))
-            x_l = moe_out + x_l  # (bs, in_features, 1)
+        if self.use_dnn and self.mlp is not None:
+            if self.architecture == "parallel":
+                deep_out = self.mlp(input_flat)
+                combined = torch.cat([cross_out, deep_out], dim=-1)
+            else:  # stacked
+                deep_out = self.mlp(cross_out)
+                combined = deep_out
+        else:
+            combined = cross_out
 
-        x_l = x_l.squeeze()  # (bs, in_features)
-        return x_l
+        logit = self.final_layer(combined)
+        return self.prediction_layer(logit)

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 +59,34 @@ 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] | None = None,
+        sparse_features: list[SparseFeature] | None = None,
+        sequence_features: list[SequenceFeature] | None = None,
+        mlp_params: dict | None = None,
+        target: list[str] | 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,
+    ):
+
+        dense_features = dense_features or []
+        sparse_features = sparse_features or []
+        sequence_features = sequence_features or []
+        mlp_params = mlp_params or {}
+        optimizer_params = optimizer_params or {}
+        if loss is None:
+            loss = "bce"
+
         super(DeepFM, self).__init__(
             dense_features=dense_features,
             sparse_features=sparse_features,
@@ -87,13 +98,10 @@ 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 +115,15 @@ 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)