Help protect the Great Barrier Reef with TensorFlow on Kaggle Join Challenge

Recommending movies: ranking

View on Run in Google Colab View source on GitHub Download notebook

Real-world recommender systems are often composed of two stages:

  1. The retrieval stage is responsible for selecting an initial set of hundreds of candidates from all possible candidates. The main objective of this model is to efficiently weed out all candidates that the user is not interested in. Because the retrieval model may be dealing with millions of candidates, it has to be computationally efficient.
  2. The ranking stage takes the outputs of the retrieval model and fine-tunes them to select the best possible handful of recommendations. Its task is to narrow down the set of items the user may be interested in to a shortlist of likely candidates.

We're going to focus on the second stage, ranking. If you are interested in the retrieval stage, have a look at our retrieval tutorial.

In this tutorial, we're going to:

  1. Get our data and split it into a training and test set.
  2. Implement a ranking model.
  3. Fit and evaluate it.


Let's first get our imports out of the way.

pip install -q tensorflow-recommenders
pip install -q --upgrade tensorflow-datasets
import os
import pprint
import tempfile

from typing import Dict, Text

import numpy as np
import tensorflow as tf
import tensorflow_datasets as tfds
import tensorflow_recommenders as tfrs

Preparing the dataset

We're going to use the same data as the retrieval tutorial. This time, we're also going to keep the ratings: these are the objectives we are trying to predict.

ratings = tfds.load("movielens/100k-ratings", split="train")

ratings = x: {
    "movie_title": x["movie_title"],
    "user_id": x["user_id"],
    "user_rating": x["user_rating"]
2021-10-02 11:04:25.388548: E tensorflow/stream_executor/cuda/] failed call to cuInit: CUDA_ERROR_NO_DEVICE: no CUDA-capable device is detected

As before, we'll split the data by putting 80% of the ratings in the train set, and 20% in the test set.

shuffled = ratings.shuffle(100_000, seed=42, reshuffle_each_iteration=False)

train = shuffled.take(80_000)
test = shuffled.skip(80_000).take(20_000)

Let's also figure out unique user ids and movie titles present in the data.

This is important because we need to be able to map the raw values of our categorical features to embedding vectors in our models. To do that, we need a vocabulary that maps a raw feature value to an integer in a contiguous range: this allows us to look up the corresponding embeddings in our embedding tables.

movie_titles = ratings.batch(1_000_000).map(lambda x: x["movie_title"])
user_ids = ratings.batch(1_000_000).map(lambda x: x["user_id"])

unique_movie_titles = np.unique(np.concatenate(list(movie_titles)))
unique_user_ids = np.unique(np.concatenate(list(user_ids)))

Implementing a model


Ranking models do not face the same efficiency constraints as retrieval models do, and so we have a little bit more freedom in our choice of architectures.

A model composed of multiple stacked dense layers is a relatively common architecture for ranking tasks. We can implement it as follows:

class RankingModel(tf.keras.Model):

  def __init__(self):
    embedding_dimension = 32

    # Compute embeddings for users.
    self.user_embeddings = tf.keras.Sequential([
        vocabulary=unique_user_ids, mask_token=None),
      tf.keras.layers.Embedding(len(unique_user_ids) + 1, embedding_dimension)

    # Compute embeddings for movies.
    self.movie_embeddings = tf.keras.Sequential([
        vocabulary=unique_movie_titles, mask_token=None),
      tf.keras.layers.Embedding(len(unique_movie_titles) + 1, embedding_dimension)

    # Compute predictions.
    self.ratings = tf.keras.Sequential([
      # Learn multiple dense layers.
      tf.keras.layers.Dense(256, activation="relu"),
      tf.keras.layers.Dense(64, activation="relu"),
      # Make rating predictions in the final layer.

  def call(self, inputs):

    user_id, movie_title = inputs

    user_embedding = self.user_embeddings(user_id)
    movie_embedding = self.movie_embeddings(movie_title)

    return self.ratings(tf.concat([user_embedding, movie_embedding], axis=1))

This model takes user ids and movie titles, and outputs a predicted rating:

RankingModel()((["42"], ["One Flew Over the Cuckoo's Nest (1975)"]))
WARNING:tensorflow:Layers in a Sequential model should only have a single input tensor, but we receive a <class 'list'> input: ['42']
Consider rewriting this model with the Functional API.
WARNING:tensorflow:Layers in a Sequential model should only have a single input tensor, but we receive a <class 'list'> input: ['42']
Consider rewriting this model with the Functional API.
WARNING:tensorflow:Layers in a Sequential model should only have a single input tensor, but we receive a <class 'list'> input: ["One Flew Over the Cuckoo's Nest (1975)"]
Consider rewriting this model with the Functional API.
WARNING:tensorflow:Layers in a Sequential model should only have a single input tensor, but we receive a <class 'list'> input: ["One Flew Over the Cuckoo's Nest (1975)"]
Consider rewriting this model with the Functional API.
<tf.Tensor: shape=(1, 1), dtype=float32, numpy=array([[0.03740937]], dtype=float32)>

Loss and metrics

The next component is the loss used to train our model. TFRS has several loss layers and tasks to make this easy.

In this instance, we'll make use of the Ranking task object: a convenience wrapper that bundles together the loss function and metric computation.

We'll use it together with the MeanSquaredError Keras loss in order to predict the ratings.

task = tfrs.tasks.Ranking(
  loss = tf.keras.losses.MeanSquaredError(),

The task itself is a Keras layer that takes true and predicted as arguments, and returns the computed loss. We'll use that to implement the model's training loop.

The full model

We can now put it all together into a model. TFRS exposes a base model class (tfrs.models.Model) which streamlines bulding models: all we need to do is to set up the components in the __init__ method, and implement the compute_loss method, taking in the raw features and returning a loss value.

The base model will then take care of creating the appropriate training loop to fit our model.

class MovielensModel(tfrs.models.Model):

  def __init__(self):
    self.ranking_model: tf.keras.Model = RankingModel()
    self.task: tf.keras.layers.Layer = tfrs.tasks.Ranking(
      loss = tf.keras.losses.MeanSquaredError(),

  def call(self, features: Dict[str, tf.Tensor]) -> tf.Tensor:
    return self.ranking_model(
        (features["user_id"], features["movie_title"]))

  def compute_loss(self, features: Dict[Text, tf.Tensor], training=False) -> tf.Tensor:
    labels = features.pop("user_rating")

    rating_predictions = self(features)

    # The task computes the loss and the metrics.
    return self.task(labels=labels, predictions=rating_predictions)

Fitting and evaluating

After defining the model, we can use standard Keras fitting and evaluation routines to fit and evaluate the model.

Let's first instantiate the model.

model = MovielensModel()

Then shuffle, batch, and cache the training and evaluation data.

cached_train = train.shuffle(100_000).batch(8192).cache()
cached_test = test.batch(4096).cache()

Then train the model:, epochs=3)
Epoch 1/3
10/10 [==============================] - 2s 26ms/step - root_mean_squared_error: 2.1718 - loss: 4.3303 - regularization_loss: 0.0000e+00 - total_loss: 4.3303
Epoch 2/3
10/10 [==============================] - 0s 8ms/step - root_mean_squared_error: 1.1227 - loss: 1.2602 - regularization_loss: 0.0000e+00 - total_loss: 1.2602
Epoch 3/3
10/10 [==============================] - 0s 8ms/step - root_mean_squared_error: 1.1162 - loss: 1.2456 - regularization_loss: 0.0000e+00 - total_loss: 1.2456
<keras.callbacks.History at 0x7f28389eaa90>

As the model trains, the loss is falling and the RMSE metric is improving.

Finally, we can evaluate our model on the test set:

model.evaluate(cached_test, return_dict=True)
5/5 [==============================] - 2s 14ms/step - root_mean_squared_error: 1.1108 - loss: 1.2287 - regularization_loss: 0.0000e+00 - total_loss: 1.2287
{'root_mean_squared_error': 1.1108061075210571,
 'loss': 1.2062578201293945,
 'regularization_loss': 0,
 'total_loss': 1.2062578201293945}

The lower the RMSE metric, the more accurate our model is at predicting ratings.

Testing the ranking model

Now we can test the ranking model by computing predictions for a set of movies and then rank these movies based on the predictions:

test_ratings = {}
test_movie_titles = ["M*A*S*H (1970)", "Dances with Wolves (1990)", "Speed (1994)"]
for movie_title in test_movie_titles:
  test_ratings[movie_title] = model({
      "user_id": np.array(["42"]),
      "movie_title": np.array([movie_title])

for title, score in sorted(test_ratings.items(), key=lambda x: x[1], reverse=True):
  print(f"{title}: {score}")
M*A*S*H (1970): [[3.584712]]
Dances with Wolves (1990): [[3.551556]]
Speed (1994): [[3.5215874]]

Exporting for serving

The model can be easily exported for serving:, "export")
2021-10-02 11:04:38.235611: W tensorflow/python/util/] Sets are not currently considered sequences, but this may change in the future, so consider avoiding using them.
WARNING:absl:Found untraced functions such as ranking_1_layer_call_and_return_conditional_losses, ranking_1_layer_call_fn, ranking_1_layer_call_fn, ranking_1_layer_call_and_return_conditional_losses, ranking_1_layer_call_and_return_conditional_losses while saving (showing 5 of 5). These functions will not be directly callable after loading.
INFO:tensorflow:Assets written to: export/assets
INFO:tensorflow:Assets written to: export/assets

We can now load it back and perform predictions:

loaded = tf.saved_model.load("export")

loaded({"user_id": np.array(["42"]), "movie_title": ["Speed (1994)"]}).numpy()
array([[3.5215874]], dtype=float32)

Next steps

The model above gives us a decent start towards building a ranking system.

Of course, making a practical ranking system requires much more effort.

In most cases, a ranking model can be substantially improved by using more features rather than just user and candidate identifiers. To see how to do that, have a look at the side features tutorial.

A careful understanding of the objectives worth optimizing is also necessary. To get started on building a recommender that optimizes multiple objectives, have a look at our multitask tutorial.