Getting Started

Basic usage

Surprise has a set of built-in algorithms and datasets for you to play with. In its simplest form, it takes about four lines of code to evaluate the performance of an algorithm:

From file examples/
from surprise import SVD
from surprise import Dataset
from surprise import evaluate, print_perf

# Load the movielens-100k dataset (download it if needed),
# and split it into 3 folds for cross-validation.
data = Dataset.load_builtin('ml-100k')

# We'll use the famous SVD algorithm.
algo = SVD()

# Evaluate performances of our algorithm on the dataset.
perf = evaluate(algo, data, measures=['RMSE', 'MAE'])


If Surprise cannot find the movielens-100k dataset, it will offer to download it and will store it under the .surprise_data folder in your home directory. The split() method automatically splits the dataset into 3 folds and the evaluate() function runs the cross-validation procedure and compute some accuracy measures.

Load a custom dataset

You can of course use a custom dataset. Surprise offers two ways of loading a custom dataset:

  • you can either specify a single file with all the ratings and use the split () method to perform cross-validation ;
  • or if your dataset is already split into predefined folds, you can specify a list of files for training and testing.

Either way, you will need to define a Reader object for Surprise to be able to parse the file(s).

We’ll see how to handle both cases with the movielens-100k dataset. Of course this is a built-in dataset, but we will act as if it were not.

Load an entire dataset

From file examples/
# path to dataset file
file_path = os.path.expanduser('~/.surprise_data/ml-100k/ml-100k/')

# As we're loading a custom dataset, we need to define a reader. In the
# movielens-100k dataset, each line has the following format:
# 'user item rating timestamp', separated by '\t' characters.
reader = Reader(line_format='user item rating timestamp', sep='\t')

data = Dataset.load_from_file(file_path, reader=reader)


Actually, as the Movielens-100k dataset is builtin, Surprise provides with a proper reader so in this case, we could have just created the reader like this:

reader = Reader('ml-100k')

For more details about readers and how to use them, see the Reader class documentation.

Load a dataset with predefined folds

From file examples/
# path to dataset folder
files_dir = os.path.expanduser('~/.surprise_data/ml-100k/ml-100k/')

# This time, we'll use the built-in reader.
reader = Reader('ml-100k')

# folds_files is a list of tuples containing file paths:
# [(u1.base, u1.test), (u2.base, u2.test), ... (u5.base, u5.test)]
train_file = files_dir + 'u%d.base'
test_file = files_dir + 'u%d.test'
folds_files = [(train_file % i, test_file % i) for i in (1, 2, 3, 4, 5)]

data = Dataset.load_from_folds(folds_files, reader=reader)

Of course, nothing prevents you from only loading a single file for training and a single file for testing. However, the folds_files parameter still needs to be a list.

Advanced usage

We will here get a little deeper on what can Surprise do for you.

Tune algorithm parameters with GridSearch

The evaluate() function gives us the results on one set of parameters given to the algorithm. If the user wants to try the algorithm on a different set of parameters, the GridSearch class comes to the rescue. Given a dict of parameters, this class exhaustively tries all the combination of parameters and helps get the best combination for an accuracy measurement. It is analogous to GridSearchCV from scikit-learn.

For instance, suppose that we want to tune the parameters of the SVD. Some of the parameters of this algorithm are n_epochs, lr_all and reg_all. Thus we define a parameters grid as follows

From file examples/
param_grid = {'n_epochs': [5, 10], 'lr_all': [0.002, 0.005],
              'reg_all': [0.4, 0.6]}

Next we define a GridSearch instance and give it the class SVD as an algorithm, and param_grid. We will compute both the RMSE and FCP values for all the combination. Thus the following definition:

From file examples/
grid_search = GridSearch(SVD, param_grid, measures=['RMSE', 'FCP'])

Now that GridSearch instance is ready, we can evaluate the algorithm on any data with the GridSearch.evaluate() method, exactly like with the regular evaluate() function:

From file examples/
data = Dataset.load_builtin('ml-100k')


Everything is ready now to read the results. For example, we get the best RMSE and FCP scores and parameters as follows:

From file examples/
# best RMSE score
# >>> 0.96117566386

# combination of parameters that gave the best RMSE score
# >>> {'reg_all': 0.4, 'lr_all': 0.005, 'n_epochs': 10}

# best FCP score
# >>> 0.702279736531

# combination of parameters that gave the best FCP score
# >>> {'reg_all': 0.6, 'lr_all': 0.005, 'n_epochs': 10}

For further analysis, we can easily read all the results in a pandas DataFrame as follows:

From file examples/
import pandas as pd  # noqa

results_df = pd.DataFrame.from_dict(grid_search.cv_results)

Manually iterate over folds

We have so far used the evaluate() function that does all the hard work for us. If you want to have better control on your experiments, you can use the folds() generator of your dataset, and then the train() and test() methods of your algorithm on each of the folds:

From file examples/
data = Dataset.load_builtin('ml-100k')

algo = BaselineOnly()

for trainset, testset in data.folds():

    # train and test algorithm.
    predictions = algo.test(testset)

    # Compute and print Root Mean Squared Error
    rmse = accuracy.rmse(predictions, verbose=True)

Train on a whole trainset and specifically query for predictions

We will here review how to get a prediction for specified users and items. In the mean time, we will also review how to train on a whole dataset, without performing cross-validation (i.e. there is no test set).

The latter is pretty straightforward: all you need is to load a dataset, and the build_full_trainset() method to build the trainset and train you algorithm:

From file examples/
data = Dataset.load_builtin('ml-100k')

# Retrieve the trainset.
trainset = data.build_full_trainset()

# Build an algorithm, and train it.
algo = KNNBasic()

Now, there’s no way we could call the test() method, because we have no testset. But you can still get predictions for the users and items you want.

Let’s say you’re interested in user 196 and item 302 (make sure they’re in the trainset!), and you know that the true rating \(r_{ui} = 4\). All you need is call the predict() method:

From file examples/
uid = str(196)  # raw user id (as in the ratings file). They are **strings**!
iid = str(302)  # raw item id (as in the ratings file). They are **strings**!

# get a prediction for specific users and items.
pred = algo.predict(uid, iid, r_ui=4, verbose=True)

If the predict() method is called with user or item ids that were not part of the trainset, it’s up to the algorithm to decide if it still can make a prediction or not. If it can’t, predict() will still predict the mean of all ratings \(\mu\).


Raw ids are ids as defined in a rating file. They can be strings, numbers, or whatever (but are still represented as strings). On trainset creation, each raw id is mapped to a unique integer called inner id, which is a lot more suitable for Surprise to manipulate. Conversions between raw and inner ids can be done using the to_inner_uid(), to_inner_iid(), to_raw_uid(), and to_raw_iid() methods of the trainset.

Obviously, it is perfectly fine to use the predict() method directly during a cross-validation process. It’s then up to you to ensure that the user and item ids are present in the trainset though.

Command line usage

Surprise can also be used from the command line, for example:

surprise -algo SVD -params "{'n_epochs': 5, 'verbose': True}" -load-builtin ml-100k -n-folds 3

See detailed usage by running:

surprise -h