A Comprehensive Overview of Sequence Recommendation Models and Techniques

The article introduces the sequential recommendation problem, defining user behavior sequences and emphasizing the importance of temporal order in predicting future items.

Data preparation steps are discussed, such as filtering invalid users, handling popular items, splitting data without future leakage, and incorporating contextual features.

MLP Models : Describes how multi‑layer perceptrons embed recent user actions, pool them, and use fully connected layers for recall and ranking, with variations like attention pooling to weight historical actions.

CNN Models : Explains using convolutional neural networks to capture local patterns in user behavior, combining global and local information, and provides a pseudo‑code example:

pooling_outputs = []
for kernel_size in kernels:
    # 1‑D convolution equivalent to [kernel_size, input_emb_size, 1, output_dimension]
    conv_out = tf.layers.conv1d(input_emb, kernel_size, output_dimension, strides=1, padding='VALID')
    # Max‑pool over the sequence dimension
    pool_out = tf.reduce_max(conv_out, -2)
    pooling_outputs.append(pool_out)
return tf.concat(pooling_outputs, -1)

RNN Models : Covers GRU‑based session models, integrating item IDs with features, and shows a TensorFlow implementation for causal convolution with padding:

def conv_with_padding(input_sequence, kernel_size, dilation_rate, output_filters):
    padding_size = (kernel_size - 1) * dilation_rate
    padded_sequence = tf.pad(input_sequence, [[0,0], [padding_size,0], [0,0]])
    conv_output = tf.layers.conv1d(padded_sequence, filters=output_filters,
                                  kernel_size=kernel_size, activation=None,
                                  padding='VALID', strides=1,
                                  dilated_rate=dilation_rate)
    return conv_output

Temporal CNN (TCN) is introduced with causal, dilated convolutions and residual connections to model long sequences efficiently.

Self‑Attention (Transformer) models are presented, detailing position embeddings, multi‑head attention, feed‑forward layers, layer normalization, and residual connections, highlighting their ability to capture global dependencies.

Reinforcement Learning is discussed as an alternative view, modeling recommendation as a Markov Decision Process with states, actions, rewards, transition probabilities, discount factors, policies, and value functions, and outlines policy‑gradient optimization.

The article concludes with practical advice on model selection—favoring simple, well‑supported models first—and provides an extensive reference list.