Implementation:Recommenders team Recommenders Caser Model

Knowledge Sources	Recommenders
Domains	Recommendation Systems, Deep Learning, Sequential Recommendation
Last Updated	2026-02-10 00:00 GMT

Overview

Implements the Caser (Convolutional Sequence Embedding) model for top-N sequential recommendation using both vertical and horizontal convolution operations over user interaction history.

Description

The CaserModel class extends SequentialBaseModel and uses CNN-based sequence modeling. The constructor initializes key hyperparameters: L (history sequence length for convolution), T (prediction shape), n_v (number of vertical convolution filters), n_h (number of horizontal convolution filters), and lengths (filter widths from 1 to L for horizontal convolutions).

The _build_seq_graph method calls _caser_cnn, which applies two parallel CNN branches: one for item_history_embedding and one for cate_history_embedding. Each branch uses the _add_cnn method which consists of:

• Vertical convolutions: The embedding matrix is transposed and processed with n_v filters of full-embedding-width via _build_cnn, then flattened. This captures point-level sequential patterns across the entire feature dimension.
• Horizontal convolutions: Multiple convolutions with filter widths ranging from 1 to L, each using n_h filters with ReLU activation via tf.compat.v1.layers.conv1d. Max-pooling over the time dimension is applied to each output, and all pooled results are concatenated. This captures union-level sequential patterns of varying lengths.

The vertical and horizontal outputs are concatenated, and the item and category CNN outputs are further concatenated into cnn_output. The final model output combines this with target_item_embedding.

Usage

Use this model when you want a CNN-based approach to sequential recommendation that captures both individual item signals (vertical) and multi-item sequential patterns (horizontal). It provides an alternative to RNN-based approaches like GRU for sequence modeling.

Code Reference

Source Location

• Repository: Recommenders
• File: recommenders/models/deeprec/models/sequential/caser.py
• Lines: 1-108

Signature

class CaserModel(SequentialBaseModel):
    def __init__(self, hparams, iterator_creator, seed=None)

    def _build_seq_graph(self)

    def _add_cnn(self, hist_matrix, vertical_dim, scope)

    def _caser_cnn(self)

    def _build_cnn(self, history_matrix, nums, shape)

Import

from recommenders.models.deeprec.models.sequential.caser import CaserModel

I/O Contract

Inputs

Name	Type	Required	Description
hparams	HParams	Yes	Hyperparameters object containing L, T, n_v, n_h, and other model configuration
iterator_creator	object	Yes	An iterator class for loading sequential data
seed	int	No	Random seed for reproducibility
hist_matrix (_add_cnn)	tf.Tensor	Yes	History sequential embedding tensor
vertical_dim (_add_cnn)	int	Yes	Embedding dimension for vertical convolution
scope (_add_cnn)	str	Yes	Variable scope name for the CNN operations

Outputs

Name	Type	Description
return (_build_seq_graph)	tf.Tensor	Concatenation of CNN-encoded sequence representation and target item embedding
return (_add_cnn)	tf.Tensor	Concatenated vertical and horizontal convolution outputs
return (_caser_cnn)	tf.Tensor	Concatenated item and category CNN outputs
return (_build_cnn)	tf.Tensor	Output of a single 1D convolution layer with ReLU activation

Usage Examples

Basic Usage

from recommenders.models.deeprec.models.sequential.caser import CaserModel
from recommenders.models.deeprec.deeprec_utils import prepare_hparams

# Prepare hyperparameters from YAML config
hparams = prepare_hparams(
    "recommenders/models/deeprec/config/caser.yaml",
    L=5,      # history sequence length
    T=3,      # prediction length
    n_v=4,    # number of vertical filters
    n_h=16,   # number of horizontal filters
)

# Create and train the Caser model
model = CaserModel(hparams, iterator_creator)
model.fit(train_file, valid_file)

# Evaluate the model
eval_results = model.run_eval(test_file)