Source code for torch_geometric.nn.models.lightgcn

from typing import Optional, Union

import torch
import torch.nn.functional as F
from torch import Tensor
from torch.nn import Embedding, ModuleList
from torch.nn.modules.loss import _Loss

from torch_geometric.nn.conv import LGConv
from torch_geometric.typing import Adj, OptTensor
from torch_geometric.utils import is_sparse, to_edge_index


[docs]class LightGCN(torch.nn.Module): r"""The LightGCN model from the `"LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation" <https://arxiv.org/abs/2002.02126>`_ paper. :class:`~torch_geometric.nn.models.LightGCN` learns embeddings by linearly propagating them on the underlying graph, and uses the weighted sum of the embeddings learned at all layers as the final embedding .. math:: \textbf{x}_i = \sum_{l=0}^{L} \alpha_l \textbf{x}^{(l)}_i, where each layer's embedding is computed as .. math:: \mathbf{x}^{(l+1)}_i = \sum_{j \in \mathcal{N}(i)} \frac{1}{\sqrt{\deg(i)\deg(j)}}\mathbf{x}^{(l)}_j. Two prediction heads and training objectives are provided: **link prediction** (via :meth:`~torch_geometric.nn.models.LightGCN.link_pred_loss` and :meth:`~torch_geometric.nn.models.LightGCN.predict_link`) and **recommendation** (via :meth:`~torch_geometric.nn.models.LightGCN.recommendation_loss` and :meth:`~torch_geometric.nn.models.LightGCN.recommend`). .. note:: Embeddings are propagated according to the graph connectivity specified by :obj:`edge_index` while rankings or link probabilities are computed according to the edges specified by :obj:`edge_label_index`. .. note:: For an example of using :class:`LightGCN`, see `examples/lightgcn.py <https://github.com/pyg-team/pytorch_geometric/blob/master/examples/ lightgcn.py>`_. Args: num_nodes (int): The number of nodes in the graph. embedding_dim (int): The dimensionality of node embeddings. num_layers (int): The number of :class:`~torch_geometric.nn.conv.LGConv` layers. alpha (float or torch.Tensor, optional): The scalar or vector specifying the re-weighting coefficients for aggregating the final embedding. If set to :obj:`None`, the uniform initialization of :obj:`1 / (num_layers + 1)` is used. (default: :obj:`None`) **kwargs (optional): Additional arguments of the underlying :class:`~torch_geometric.nn.conv.LGConv` layers. """ def __init__( self, num_nodes: int, embedding_dim: int, num_layers: int, alpha: Optional[Union[float, Tensor]] = None, **kwargs, ): super().__init__() self.num_nodes = num_nodes self.embedding_dim = embedding_dim self.num_layers = num_layers if alpha is None: alpha = 1. / (num_layers + 1) if isinstance(alpha, Tensor): assert alpha.size(0) == num_layers + 1 else: alpha = torch.tensor([alpha] * (num_layers + 1)) self.register_buffer('alpha', alpha) self.embedding = Embedding(num_nodes, embedding_dim) self.convs = ModuleList([LGConv(**kwargs) for _ in range(num_layers)]) self.reset_parameters()
[docs] def reset_parameters(self): r"""Resets all learnable parameters of the module.""" torch.nn.init.xavier_uniform_(self.embedding.weight) for conv in self.convs: conv.reset_parameters()
[docs] def get_embedding( self, edge_index: Adj, edge_weight: OptTensor = None, ) -> Tensor: r"""Returns the embedding of nodes in the graph.""" x = self.embedding.weight out = x * self.alpha[0] for i in range(self.num_layers): x = self.convs[i](x, edge_index, edge_weight) out = out + x * self.alpha[i + 1] return out
[docs] def forward( self, edge_index: Adj, edge_label_index: OptTensor = None, edge_weight: OptTensor = None, ) -> Tensor: r"""Computes rankings for pairs of nodes. Args: edge_index (torch.Tensor or SparseTensor): Edge tensor specifying the connectivity of the graph. edge_label_index (torch.Tensor, optional): Edge tensor specifying the node pairs for which to compute rankings or probabilities. If :obj:`edge_label_index` is set to :obj:`None`, all edges in :obj:`edge_index` will be used instead. (default: :obj:`None`) edge_weight (torch.Tensor, optional): The weight of each edge in :obj:`edge_index`. (default: :obj:`None`) """ if edge_label_index is None: if is_sparse(edge_index): edge_label_index, _ = to_edge_index(edge_index) else: edge_label_index = edge_index out = self.get_embedding(edge_index, edge_weight) out_src = out[edge_label_index[0]] out_dst = out[edge_label_index[1]] return (out_src * out_dst).sum(dim=-1)
[docs] def recommend( self, edge_index: Adj, edge_weight: OptTensor = None, src_index: OptTensor = None, dst_index: OptTensor = None, k: int = 1, sorted: bool = True, ) -> Tensor: r"""Get top-:math:`k` recommendations for nodes in :obj:`src_index`. Args: edge_index (torch.Tensor or SparseTensor): Edge tensor specifying the connectivity of the graph. edge_weight (torch.Tensor, optional): The weight of each edge in :obj:`edge_index`. (default: :obj:`None`) src_index (torch.Tensor, optional): Node indices for which recommendations should be generated. If set to :obj:`None`, all nodes will be used. (default: :obj:`None`) dst_index (torch.Tensor, optional): Node indices which represent the possible recommendation choices. If set to :obj:`None`, all nodes will be used. (default: :obj:`None`) k (int, optional): Number of recommendations. (default: :obj:`1`) sorted (bool, optional): Whether to sort the recommendations by score. (default: :obj:`True`) """ out_src = out_dst = self.get_embedding(edge_index, edge_weight) if src_index is not None: out_src = out_src[src_index] if dst_index is not None: out_dst = out_dst[dst_index] pred = out_src @ out_dst.t() top_index = pred.topk(k, dim=-1, sorted=sorted).indices if dst_index is not None: # Map local top-indices to original indices. top_index = dst_index[top_index.view(-1)].view(*top_index.size()) return top_index
[docs] def recommendation_loss( self, pos_edge_rank: Tensor, neg_edge_rank: Tensor, node_id: Optional[Tensor] = None, lambda_reg: float = 1e-4, **kwargs, ) -> Tensor: r"""Computes the model loss for a ranking objective via the Bayesian Personalized Ranking (BPR) loss. .. note:: The i-th entry in the :obj:`pos_edge_rank` vector and i-th entry in the :obj:`neg_edge_rank` entry must correspond to ranks of positive and negative edges of the same entity (*e.g.*, user). Args: pos_edge_rank (torch.Tensor): Positive edge rankings. neg_edge_rank (torch.Tensor): Negative edge rankings. node_id (torch.Tensor): The indices of the nodes involved for deriving a prediction for both positive and negative edges. If set to :obj:`None`, all nodes will be used. lambda_reg (int, optional): The :math:`L_2` regularization strength of the Bayesian Personalized Ranking (BPR) loss. (default: :obj:`1e-4`) **kwargs (optional): Additional arguments of the underlying :class:`torch_geometric.nn.models.lightgcn.BPRLoss` loss function. """ loss_fn = BPRLoss(lambda_reg, **kwargs) emb = self.embedding.weight emb = emb if node_id is None else emb[node_id] return loss_fn(pos_edge_rank, neg_edge_rank, emb)
def __repr__(self) -> str: return (f'{self.__class__.__name__}({self.num_nodes}, ' f'{self.embedding_dim}, num_layers={self.num_layers})')
class BPRLoss(_Loss): r"""The Bayesian Personalized Ranking (BPR) loss. The BPR loss is a pairwise loss that encourages the prediction of an observed entry to be higher than its unobserved counterparts (see `here <https://arxiv.org/abs/2002.02126>`__). .. math:: L_{\text{BPR}} = - \sum_{u=1}^{M} \sum_{i \in \mathcal{N}_u} \sum_{j \not\in \mathcal{N}_u} \ln \sigma(\hat{y}_{ui} - \hat{y}_{uj}) + \lambda \vert\vert \textbf{x}^{(0)} \vert\vert^2 where :math:`\lambda` controls the :math:`L_2` regularization strength. We compute the mean BPR loss for simplicity. Args: lambda_reg (float, optional): The :math:`L_2` regularization strength (default: 0). **kwargs (optional): Additional arguments of the underlying :class:`torch.nn.modules.loss._Loss` class. """ __constants__ = ['lambda_reg'] lambda_reg: float def __init__(self, lambda_reg: float = 0, **kwargs): super().__init__(None, None, "sum", **kwargs) self.lambda_reg = lambda_reg def forward(self, positives: Tensor, negatives: Tensor, parameters: Tensor = None) -> Tensor: r"""Compute the mean Bayesian Personalized Ranking (BPR) loss. .. note:: The i-th entry in the :obj:`positives` vector and i-th entry in the :obj:`negatives` entry should correspond to the same entity (*.e.g*, user), as the BPR is a personalized ranking loss. Args: positives (Tensor): The vector of positive-pair rankings. negatives (Tensor): The vector of negative-pair rankings. parameters (Tensor, optional): The tensor of parameters which should be used for :math:`L_2` regularization (default: :obj:`None`). """ log_prob = F.logsigmoid(positives - negatives).mean() regularization = 0 if self.lambda_reg != 0: regularization = self.lambda_reg * parameters.norm(p=2).pow(2) regularization = regularization / positives.size(0) return -log_prob + regularization