Skip to content

An AutoRecSys library for Surprise. Automate algorithm selection and hyperparameter tuning 🚀

License

Notifications You must be signed in to change notification settings

ISG-Siegen/Auto-Surprise

Repository files navigation

Auto-Surprise

GitHub release (latest by date) PyPI Downloads

Auto-Surprise is built as a wrapper around the Python Surprise recommender-system library. It automates algorithm selection and hyper parameter optimization in a highly parallelized manner.

Setup

Auto-Surprise is easy to install with Pip, and required Python>=3.6 installed on a linux system. Currently not supported in windows, but can be used using WSL.

$ pip install auto-surprise

Usage

Basic usage of AutoSurprise is given below.

from surprise import Dataset
from auto_surprise.engine import Engine

# Load the dataset
data = Dataset.load_builtin('ml-100k')

# Intitialize auto surprise engine
engine = Engine(verbose=True)

# Start the trainer
best_algo, best_params, best_score, tasks = engine.train(
    data=data, 
    target_metric='test_rmse', 
    cpu_time_limit=60 * 60, 
    max_evals=100
)

In the above example, we first initialize the Engine. We then run engine.train() to begin training our model. To train the model we need to pass the following

  • data : The data as an instance of surprise.dataset.DatasetAutoFolds. Please read Surprise Dataset docs
  • target_metric : The metric we seek to minimize. Available options are test_rmse and test_mae.
  • cpu_time_limit : The time limit we want to train. This is in seconds. For datasets like Movielens 100k, 1 hour is sufficient. But you may want to increase this based on the size of your dataset
  • max_evals: The maximum number of evaluations each algorithm gets for hyper parameter optimization.
  • hpo_algo: Auto-Surprise uses Hyperopt for hyperparameter tuning. By default, it's set to use TPE, but you can change this to any algorithm supported by hyperopt, such as Adaptive TPE or Random search.

Setting the Hyperparameter Optimization Algorithm

Auto-Surprise uses Hyperopt. You can change the HPO algo as shown below.

# Example for setting the HPO algorithm to adaptive TPE
import hyperopt

...

engine = Engine(verbose=True)
engine.train(
    data=data,
    target_metric='test_rmse',
    cpu_time_limit=60 * 60,
    max_evals=100,
    hpo_algo=hyperopt.atpe.suggest
)

Building back the best model

You can build a pickelable model as shown.

model = engine.build_model(best_algo, best_params)

Benchmarks

In my testing, Auto-Surprise performed anywhere from 0.8 to 4% improvement in RMSE compared to the best performing default algorithm configuration. In the table below are the results for the Jester 2 dataset. Benchmark results for Movielens and Book-Crossing dataset are also available here

Algorithm RMSE MAE Time
Normal Predictor 7.277 5.886 00:00:01
SVD 4.905 3.97 00:00:13
SVD++ 5.102 4.055 00:00:29
NMF -- -- --
Slope One 5.189 3.945 00:00:02
KNN Basic 5.078 4.034 00:02:14
KNN with Means 5.124 3.955 00:02:16
KNN with Z-score 5.219 3.955 00:02:20
KNN Baseline 4.898 3.896 00:02:14
Co-clustering 5.153 3.917 00:00:12
Baseline Only 4.849 3.934 00:00:01
GridSearch 4.7409 3.8147 80:52:35
Auto-Surprise (TPE) 4.6489 3.6837 02:00:10
Auto-Surprise (ATPE) 4.6555 3.6906 02:00:01

Papers

Auto-Surprise: An Automated Recommender-System (AutoRecSys) Library with Tree of Parzens Estimator (TPE) Optimization