Code examples / Structured Data / A Transformer-based recommendation system

A Transformer-based recommendation system

Author: Khalid Salama
Date created: 2020/12/30
Last modified: 2020/12/30
Description: Rating rate prediction using the Behavior Sequence Transformer (BST) model on the Movielens.

ⓘ This example uses Keras 3

View in Colab GitHub source


This example demonstrates the Behavior Sequence Transformer (BST) model, by Qiwei Chen et al., using the Movielens dataset. The BST model leverages the sequential behaviour of the users in watching and rating movies, as well as user profile and movie features, to predict the rating of the user to a target movie.

More precisely, the BST model aims to predict the rating of a target movie by accepting the following inputs:

  1. A fixed-length sequence of movie_ids watched by a user.
  2. A fixed-length sequence of the ratings for the movies watched by a user.
  3. A set of user features, including user_id, sex, occupation, and age_group.
  4. A set of genres for each movie in the input sequence and the target movie.
  5. A target_movie_id for which to predict the rating.

This example modifies the original BST model in the following ways:

  1. We incorporate the movie features (genres) into the processing of the embedding of each movie of the input sequence and the target movie, rather than treating them as "other features" outside the transformer layer.
  2. We utilize the ratings of movies in the input sequence, along with the their positions in the sequence, to update them before feeding them into the self-attention layer.

Note that this example should be run with TensorFlow 2.4 or higher.

The dataset

We use the 1M version of the Movielens dataset. The dataset includes around 1 million ratings from 6000 users on 4000 movies, along with some user features, movie genres. In addition, the timestamp of each user-movie rating is provided, which allows creating sequences of movie ratings for each user, as expected by the BST model.


import os

os.environ["KERAS_BACKEND"] = "tensorflow"

import math
from zipfile import ZipFile
from urllib.request import urlretrieve

import keras
import numpy as np
import pandas as pd
import tensorflow as tf
from keras import layers
from keras.layers import StringLookup

Prepare the data

Download and prepare the DataFrames

First, let's download the movielens data.

The downloaded folder will contain three data files: users.dat, movies.dat, and ratings.dat.

urlretrieve("", "")
ZipFile("", "r").extractall()

Then, we load the data into pandas DataFrames with their proper column names.

users = pd.read_csv(
    names=["user_id", "sex", "age_group", "occupation", "zip_code"],

ratings = pd.read_csv(
    names=["user_id", "movie_id", "rating", "unix_timestamp"],

movies = pd.read_csv(
    names=["movie_id", "title", "genres"],

Here, we do some simple data processing to fix the data types of the columns.

users["user_id"] = users["user_id"].apply(lambda x: f"user_{x}")
users["age_group"] = users["age_group"].apply(lambda x: f"group_{x}")
users["occupation"] = users["occupation"].apply(lambda x: f"occupation_{x}")

movies["movie_id"] = movies["movie_id"].apply(lambda x: f"movie_{x}")

ratings["movie_id"] = ratings["movie_id"].apply(lambda x: f"movie_{x}")
ratings["user_id"] = ratings["user_id"].apply(lambda x: f"user_{x}")
ratings["rating"] = ratings["rating"].apply(lambda x: float(x))

Each movie has multiple genres. We split them into separate columns in the movies DataFrame.

genres = ["Action", "Adventure", "Animation", "Children's", "Comedy", "Crime"]
genres += ["Documentary", "Drama", "Fantasy", "Film-Noir", "Horror", "Musical"]
genres += ["Mystery", "Romance", "Sci-Fi", "Thriller", "War", "Western"]

for genre in genres:
    movies[genre] = movies["genres"].apply(
        lambda values: int(genre in values.split("|"))

Transform the movie ratings data into sequences

First, let's sort the the ratings data using the unix_timestamp, and then group the movie_id values and the rating values by user_id.

The output DataFrame will have a record for each user_id, with two ordered lists (sorted by rating datetime): the movies they have rated, and their ratings of these movies.

ratings_group = ratings.sort_values(by=["unix_timestamp"]).groupby("user_id")

ratings_data = pd.DataFrame(
        "user_id": list(ratings_group.groups.keys()),
        "movie_ids": list(ratings_group.movie_id.apply(list)),
        "ratings": list(ratings_group.rating.apply(list)),
        "timestamps": list(ratings_group.unix_timestamp.apply(list)),

Now, let's split the movie_ids list into a set of sequences of a fixed length. We do the same for the ratings. Set the sequence_length variable to change the length of the input sequence to the model. You can also change the step_size to control the number of sequences to generate for each user.

sequence_length = 4
step_size = 2

def create_sequences(values, window_size, step_size):
    sequences = []
    start_index = 0
    while True:
        end_index = start_index + window_size
        seq = values[start_index:end_index]
        if len(seq) < window_size:
            seq = values[-window_size:]
            if len(seq) == window_size:
        start_index += step_size
    return sequences

ratings_data.movie_ids = ratings_data.movie_ids.apply(
    lambda ids: create_sequences(ids, sequence_length, step_size)

ratings_data.ratings = ratings_data.ratings.apply(
    lambda ids: create_sequences(ids, sequence_length, step_size)

del ratings_data["timestamps"]

After that, we process the output to have each sequence in a separate records in the DataFrame. In addition, we join the user features with the ratings data.

ratings_data_movies = ratings_data[["user_id", "movie_ids"]].explode(
    "movie_ids", ignore_index=True
ratings_data_rating = ratings_data[["ratings"]].explode("ratings", ignore_index=True)
ratings_data_transformed = pd.concat([ratings_data_movies, ratings_data_rating], axis=1)
ratings_data_transformed = ratings_data_transformed.join(
    users.set_index("user_id"), on="user_id"
ratings_data_transformed.movie_ids = ratings_data_transformed.movie_ids.apply(
    lambda x: ",".join(x)
ratings_data_transformed.ratings = ratings_data_transformed.ratings.apply(
    lambda x: ",".join([str(v) for v in x])

del ratings_data_transformed["zip_code"]

    columns={"movie_ids": "sequence_movie_ids", "ratings": "sequence_ratings"},

With sequence_length of 4 and step_size of 2, we end up with 498,623 sequences.

Finally, we split the data into training and testing splits, with 85% and 15% of the instances, respectively, and store them to CSV files.

random_selection = np.random.rand(len(ratings_data_transformed.index)) <= 0.85
train_data = ratings_data_transformed[random_selection]
test_data = ratings_data_transformed[~random_selection]

train_data.to_csv("train_data.csv", index=False, sep="|", header=False)
test_data.to_csv("test_data.csv", index=False, sep="|", header=False)

Define metadata

CSV_HEADER = list(ratings_data_transformed.columns)

    "user_id": list(users.user_id.unique()),
    "movie_id": list(movies.movie_id.unique()),
    "sex": list(,
    "age_group": list(users.age_group.unique()),
    "occupation": list(users.occupation.unique()),

USER_FEATURES = ["sex", "age_group", "occupation"]

MOVIE_FEATURES = ["genres"]

Create for training and evaluation

def get_dataset_from_csv(csv_file_path, shuffle=False, batch_size=128):
    def process(features):
        movie_ids_string = features["sequence_movie_ids"]
        sequence_movie_ids = tf.strings.split(movie_ids_string, ",").to_tensor()

        # The last movie id in the sequence is the target movie.
        features["target_movie_id"] = sequence_movie_ids[:, -1]
        features["sequence_movie_ids"] = sequence_movie_ids[:, :-1]

        ratings_string = features["sequence_ratings"]
        sequence_ratings = tf.strings.to_number(
            tf.strings.split(ratings_string, ","), tf.dtypes.float32

        # The last rating in the sequence is the target for the model to predict.
        target = sequence_ratings[:, -1]
        features["sequence_ratings"] = sequence_ratings[:, :-1]

        return features, target

    dataset =

    return dataset

Create model inputs

def create_model_inputs():
    return {
        "user_id": keras.Input(name="user_id", shape=(1,), dtype="string"),
        "sequence_movie_ids": keras.Input(
            name="sequence_movie_ids", shape=(sequence_length - 1,), dtype="string"
        "target_movie_id": keras.Input(
            name="target_movie_id", shape=(1,), dtype="string"
        "sequence_ratings": keras.Input(
            name="sequence_ratings", shape=(sequence_length - 1,), dtype=tf.float32
        "sex": keras.Input(name="sex", shape=(1,), dtype="string"),
        "age_group": keras.Input(name="age_group", shape=(1,), dtype="string"),
        "occupation": keras.Input(name="occupation", shape=(1,), dtype="string"),

Encode input features

The encode_input_features method works as follows:

  1. Each categorical user feature is encoded using layers.Embedding, with embedding dimension equals to the square root of the vocabulary size of the feature. The embeddings of these features are concatenated to form a single input tensor.

  2. Each movie in the movie sequence and the target movie is encoded layers.Embedding, where the dimension size is the square root of the number of movies.

  3. A multi-hot genres vector for each movie is concatenated with its embedding vector, and processed using a non-linear layers.Dense to output a vector of the same movie embedding dimensions.

  4. A positional embedding is added to each movie embedding in the sequence, and then multiplied by its rating from the ratings sequence.

  5. The target movie embedding is concatenated to the sequence movie embeddings, producing a tensor with the shape of [batch size, sequence length, embedding size], as expected by the attention layer for the transformer architecture.

  6. The method returns a tuple of two elements: encoded_transformer_features and encoded_other_features.

def encode_input_features(
    encoded_transformer_features = []
    encoded_other_features = []

    other_feature_names = []
    if include_user_id:
    if include_user_features:

    ## Encode user features
    for feature_name in other_feature_names:
        # Convert the string input values into integer indices.
        vocabulary = CATEGORICAL_FEATURES_WITH_VOCABULARY[feature_name]
        idx = StringLookup(vocabulary=vocabulary, mask_token=None, num_oov_indices=0)(
        # Compute embedding dimensions
        embedding_dims = int(math.sqrt(len(vocabulary)))
        # Create an embedding layer with the specified dimensions.
        embedding_encoder = layers.Embedding(
        # Convert the index values to embedding representations.

    ## Create a single embedding vector for the user features
    if len(encoded_other_features) > 1:
        encoded_other_features = layers.concatenate(encoded_other_features)
    elif len(encoded_other_features) == 1:
        encoded_other_features = encoded_other_features[0]
        encoded_other_features = None

    ## Create a movie embedding encoder
    movie_vocabulary = CATEGORICAL_FEATURES_WITH_VOCABULARY["movie_id"]
    movie_embedding_dims = int(math.sqrt(len(movie_vocabulary)))
    # Create a lookup to convert string values to integer indices.
    movie_index_lookup = StringLookup(
    # Create an embedding layer with the specified dimensions.
    movie_embedding_encoder = layers.Embedding(
    # Create a vector lookup for movie genres.
    genre_vectors = movies[genres].to_numpy()
    movie_genres_lookup = layers.Embedding(
    # Create a processing layer for genres.
    movie_embedding_processor = layers.Dense(

    ## Define a function to encode a given movie id.
    def encode_movie(movie_id):
        # Convert the string input values into integer indices.
        movie_idx = movie_index_lookup(movie_id)
        movie_embedding = movie_embedding_encoder(movie_idx)
        encoded_movie = movie_embedding
        if include_movie_features:
            movie_genres_vector = movie_genres_lookup(movie_idx)
            encoded_movie = movie_embedding_processor(
                layers.concatenate([movie_embedding, movie_genres_vector])
        return encoded_movie

    ## Encoding target_movie_id
    target_movie_id = inputs["target_movie_id"]
    encoded_target_movie = encode_movie(target_movie_id)

    ## Encoding sequence movie_ids.
    sequence_movies_ids = inputs["sequence_movie_ids"]
    encoded_sequence_movies = encode_movie(sequence_movies_ids)
    # Create positional embedding.
    position_embedding_encoder = layers.Embedding(
    positions = tf.range(start=0, limit=sequence_length - 1, delta=1)
    encodded_positions = position_embedding_encoder(positions)
    # Retrieve sequence ratings to incorporate them into the encoding of the movie.
    sequence_ratings = inputs["sequence_ratings"]
    sequence_ratings = keras.ops.expand_dims(sequence_ratings, -1)
    # Add the positional encoding to the movie encodings and multiply them by rating.
    encoded_sequence_movies_with_poistion_and_rating = layers.Multiply()(
        [(encoded_sequence_movies + encodded_positions), sequence_ratings]

    # Construct the transformer inputs.
    for i in range(sequence_length - 1):
        feature = encoded_sequence_movies_with_poistion_and_rating[:, i, ...]
        feature = keras.ops.expand_dims(feature, 1)

    encoded_transformer_features = layers.concatenate(
        encoded_transformer_features, axis=1

    return encoded_transformer_features, encoded_other_features

Create a BST model

include_user_id = False
include_user_features = False
include_movie_features = False

hidden_units = [256, 128]
dropout_rate = 0.1
num_heads = 3

def create_model():
    inputs = create_model_inputs()
    transformer_features, other_features = encode_input_features(
        inputs, include_user_id, include_user_features, include_movie_features

    # Create a multi-headed attention layer.
    attention_output = layers.MultiHeadAttention(
        num_heads=num_heads, key_dim=transformer_features.shape[2], dropout=dropout_rate
    )(transformer_features, transformer_features)

    # Transformer block.
    attention_output = layers.Dropout(dropout_rate)(attention_output)
    x1 = layers.Add()([transformer_features, attention_output])
    x1 = layers.LayerNormalization()(x1)
    x2 = layers.LeakyReLU()(x1)
    x2 = layers.Dense(units=x2.shape[-1])(x2)
    x2 = layers.Dropout(dropout_rate)(x2)
    transformer_features = layers.Add()([x1, x2])
    transformer_features = layers.LayerNormalization()(transformer_features)
    features = layers.Flatten()(transformer_features)

    # Included the other features.
    if other_features is not None:
        features = layers.concatenate(
            [features, layers.Reshape([other_features.shape[-1]])(other_features)]

    # Fully-connected layers.
    for num_units in hidden_units:
        features = layers.Dense(num_units)(features)
        features = layers.BatchNormalization()(features)
        features = layers.LeakyReLU()(features)
        features = layers.Dropout(dropout_rate)(features)

    outputs = layers.Dense(units=1)(features)
    model = keras.Model(inputs=inputs, outputs=outputs)
    return model

model = create_model()

Run training and evaluation experiment

# Compile the model.

# Read the training data.
train_dataset = get_dataset_from_csv("train_data.csv", shuffle=True, batch_size=265)

# Fit the model with the training data., epochs=5)

# Read the test data.
test_dataset = get_dataset_from_csv("test_data.csv", batch_size=265)

# Evaluate the model on the test data.
_, rmse = model.evaluate(test_dataset, verbose=0)
print(f"Test MAE: {round(rmse, 3)}")
Epoch 1/5
 1600/1600 ━━━━━━━━━━━━━━━━━━━━ 19s 11ms/step - loss: 1.5762 - mean_absolute_error: 0.9892
Epoch 2/5
 1600/1600 ━━━━━━━━━━━━━━━━━━━━ 17s 11ms/step - loss: 1.1263 - mean_absolute_error: 0.8502
Epoch 3/5
 1600/1600 ━━━━━━━━━━━━━━━━━━━━ 17s 11ms/step - loss: 1.0885 - mean_absolute_error: 0.8361
Epoch 4/5
 1600/1600 ━━━━━━━━━━━━━━━━━━━━ 17s 11ms/step - loss: 1.0943 - mean_absolute_error: 0.8388
Epoch 5/5
 1600/1600 ━━━━━━━━━━━━━━━━━━━━ 17s 10ms/step - loss: 1.0360 - mean_absolute_error: 0.8142
Test MAE: 0.782

You should achieve a Mean Absolute Error (MAE) at or around 0.7 on the test data.


The BST model uses the Transformer layer in its architecture to capture the sequential signals underlying users’ behavior sequences for recommendation.

You can try training this model with different configurations, for example, by increasing the input sequence length and training the model for a larger number of epochs. In addition, you can try including other features like movie release year and customer zipcode, and including cross features like sex X genre.