Implementation:Recommenders team Recommenders EmbeddingDotBias Model

Knowledge Sources	Recommenders
Domains	Collaborative Filtering, Matrix Factorization, PyTorch
Last Updated	2026-02-10 00:00 GMT

Overview

The EmbeddingDotBias class implements a dot-product collaborative filtering model in PyTorch that predicts ratings using learned user and item embeddings combined with bias terms.

Description

The EmbeddingDotBias model is a fundamental matrix factorization approach for collaborative filtering. It creates four embedding layers: user weights, item weights, user biases, and item biases, all initialized with truncated normal distributions (std=0.01). The forward pass computes the prediction as dot(user_embedding, item_embedding) + user_bias + item_bias, optionally mapping the result through a sigmoid function scaled to a specified output range (y_range). A from_classes factory method allows building the model by inferring user and item counts from a class dictionary. Helper methods bias() and weight() allow extracting learned embeddings and biases for specific users or items by their original IDs, converting them through an internal _get_idx mapping that translates entity IDs to embedding matrix indices.

Usage

Use this model when building a collaborative filtering recommendation system that needs to predict explicit ratings from user-item interactions. It serves as a strong baseline for rating prediction tasks and is suitable for datasets where user and item IDs are known. The model supports both direct instantiation with known user/item counts and a factory pattern via from_classes for inferring counts from class dictionaries.

Code Reference

Source Location

• Repository: Recommenders
• File: recommenders/models/embdotbias/model.py
• Lines: 1-135

Signature

class EmbeddingDotBias(Module):
    def __init__(self, n_factors, n_users, n_items, y_range=None)
    def forward(self, x)
    @classmethod
    def from_classes(cls, n_factors, classes, user=None, item=None, y_range=None)
    def _get_idx(self, entity_ids, is_item=True)
    def bias(self, entity_ids, is_item=True)
    def weight(self, entity_ids, is_item=True)

Import

from recommenders.models.embdotbias.model import EmbeddingDotBias

I/O Contract

Inputs

Name	Type	Required	Description
n_factors	int	Yes	Number of latent factors for user and item embeddings
n_users	int	Yes	Total number of users in the dataset
n_items	int	Yes	Total number of items in the dataset
y_range	tuple	No	Output normalization range as (min, max); if None, raw scores are returned
x (forward)	torch.Tensor	Yes	Tensor of shape (batch_size, 2) with user indices in column 0 and item indices in column 1
classes (from_classes)	dict	Yes	Dictionary mapping entity names to lists of IDs for inferring user/item counts
entity_ids (bias/weight)	list	Yes	List of user or item IDs for embedding/bias lookup
is_item	bool	No	If True, fetch item embeddings/biases; if False, fetch user embeddings/biases (default True)

Outputs

Name	Type	Description
forward return	torch.Tensor	Predicted ratings for each user-item pair in the batch
from_classes return	EmbeddingDotBias	Instantiated model with classes attribute set
bias return	torch.Tensor	Bias values for the specified entities
weight return	torch.Tensor	Embedding weight vectors for the specified entities

Usage Examples

Basic Usage

from recommenders.models.embdotbias.model import EmbeddingDotBias

# Direct instantiation
model = EmbeddingDotBias(n_factors=40, n_users=1000, n_items=500, y_range=(1, 5))

# Using factory method with class dictionaries
classes = {"userID": [1, 2, 3, 4, 5], "itemID": [10, 20, 30]}
model = EmbeddingDotBias.from_classes(n_factors=40, classes=classes, y_range=(1, 5))

# Forward pass with a batch of user-item pairs
import torch
x = torch.tensor([[0, 1], [2, 0]])  # user indices, item indices
predictions = model(x)

# Extract learned biases and weights for specific items
item_biases = model.bias([10, 20], is_item=True)
item_weights = model.weight([10, 20], is_item=True)