irec 1.2.8

Reinforcement Learning Recommender Systems Framework

Introduction • Install • Quick Start

iRec is structured in three main components as usually made by classical frameworks in the RS field. These main components are:

• Environment Setting: responsible for loading, preprocessing, and splitting the dataset into train and test sets (when required) to create the task environment for the pipeline;

• Recommendation Agent: responsible for implementing the recommendation model required as an interactive algorithm that will interact with the environment;

• Experimental Evaluation: responsible for defining how the agent will interact with the environment to simulate the interactive scenario and get the logs required for a complete evaluation.

Introduction

For Python >= 3.8

Interactive Recommender Systems Framework

Main features:

• Several state-of-the-art reinforcement learning models for the recommendation scenario
• Novelty, coverage and much more different type of online metrics
• Integration with the most used datasets for evaluating recommendation systems
• Flexible configuration
• Modular and reusable design
• Contains multiple evaluation policies currently used in the literature to evaluate reinforcement learning models
• Online Learning and Reinforcement Learning models
• Metrics and metrics evaluators are awesome to evaluate recommender systems in different ways

Also, we provide a amazing application created using the iRec library, the iRec-cmdline, that can be used to setup a experiment under 5~ minutes with parallel processes, log registry and results views. The main features are:

• Powerful application to run any reinforcement learning experiment powered by MLflow
• Entire pipeline of execution is fully parallelized
• Log registry
• Results views
• Statistical test
• Extensible environment

Install

Install with pip:

pip install irec

Examples

iRec-cmdline contains a example of a application using iRec and MLflow, where different experiments can be run with easy using existing recommender systems.

Check this example of a execution using the example application:

dataset=("Netflix 10k" "Good Books" "Yahoo Music 10k");\
models=(Random MostPopular UCB ThompsonSampling EGreedy);\
metrics=(Hits Precision Recall);\
eval_pol=("FixedInteraction");
metric_evaluator="Interaction";\

cd agents &&
python run_agent_best.py --agents "${models[@]}" --dataset_loaders "${dataset[@]}" --evaluation_policy "${eval_pol[@]}" &&

cd ../evaluation &&
python eval_agent_best.py --agents "${models[@]}" --dataset_loaders "${dataset[@]}" --evaluation_policy "${eval_pol[@]}" --metrics "${metrics[@]}" --metric_evaluator "${metric_eval[@]}" &&

python print_latex_table_results.py --agents "${models[@]}" --dataset_loaders "${dataset[@]}" --evaluation_policy "${eval_pol[@]}" --metric_evaluator "${metric_eval[@]}" --metrics "${metrics[@]}"

Datasets

Our framework has the ability to use any type of dataset, as long as it is suitable for the recommendation domain and is formatted correctly. Below we list some datasets tested and used in some of our experiments.

Dataset	Domain	Sparsity	Link
MovieLens 100k	Movies	93.69%	Link
MovieLens 1M	Movies	95.80%	Link
MovieLens 10M	Movies	98.66%	Link
Netflix	Movies	98.69%	Link
Ciao DVD	Movies	99.97%	Link
Yahoo Music	Musics	97.63%	Link
LastFM	Musics	99.84%	Link
Good Books	Books	98.88%	Link
Good Reads	Books	99.50%	Link
Amazon Kindle Store	Products	99.97%	Link
Clothing Fit	Clothes	99.97%	Link

Models

The recommender models supported by irec are listed below.

Year	Model	Paper	Description
2002	ε-Greedy	Link	In general, ε-Greedy models the problem based on an ε diversification parameter to perform random actions.
2013	Linear ε-Greedy	Link	A linear exploitation of the items latent factors defined by a PMF formulation that also explore random items with probability ε.
2011	Thompson Sampling	Link	A basic item-oriented bandit algorithm that follows a Gaussian distribution of items and users to perform the prediction rule based on their samples.
2013	GLM-UCB	Link	It follows a similar process as Linear UCB based on the PMF formulation, but it also adds a sigmoid form in the exploitation step and makes a time-dependent exploration.
2018	ICTR	Link	It is an interactive collaborative topic regression model that utilizes the TS bandit algorithm and controls the items dependency by a particle learning strategy.
2015	PTS	Link	It is a PMF formulation for the original TS based on a Bayesian inference around the items. This method also applies particle filtering to guide the exploration of items over time.
2019	kNN Bandit	Link	A simple multi-armed bandit elaboration of neighbor-based collaborative filtering. A variant of the nearest-neighbors scheme, but endowed with a controlled stochastic exploration capability of the users’ neighborhood, by a parameter-free application of Thompson sampling.
2017	Linear TS	Link	An adaptation of the original Thompson Sampling to measure the latent dimensions by a PMF formulation.
2013	Linear UCB	Link	An adaptation of the original LinUCB (Lihong Li et al. 2010) to measure the latent dimensions by a PMF formulation.
2020	NICF	Link	It is an interactive method based on a combination of neural networks and collaborative filtering that also performs a meta-learning of the user’s preferences.
2016	COFIBA	Link	This method relies on upper-confidence-based tradeoffs between exploration and exploitation, combined with adaptive clustering procedures at both the user and the item sides.
2002	UCB	Link	It is the original UCB that calculates a confidence interval for each item at each iteration and tries to shrink the confidence bounds.
2021	Cluster-Bandit (CB)	Link	it is a new bandit algorithm based on clusters to face the cold-start problem.
2002	Entropy	Link	The entropy of an item i is calculated using the relative frequency of the possible ratings. In general, since entropy measures the spread of ratings for an item, this strategy tends to promote rarely rated items, which can be considerably informative.
2002	log(pop)*ent	Link	It combines popularity and entropy to identify potentially relevant items that also have the ability to add more knowledge to the system. As these concepts are not strongly correlated, it is possible to achieve this combination through a linear combination of the popularity ρ of an item i by its entropy ε: score(i) = log(ρi) · εi.
-	Random	Link	This method recommends totally random items.
-	Most Popular	Link	It recommends items with the higher number of ratings received (most-popular) at each iteration.
-	Best Rated	Link	Recommends top-rated items based on their average ratings in each iteration.

Metrics

The recommender metrics supported by iRec are listed below.

Metric	Reference	Description
Hits	Link	Number of recommendations made successfully.
Precision	Link	Precision is defined as the percentage of predictions we get right.
Recall	Link	Represents the probability that a relevant item will be selected.
EPC	Link	Represents the novelty for each user and it is measured by the expected number of seen relevant recommended items not previously seen.
EPD	Link	EPD is a distance-based novelty measure, which looks at distances between the items in the user’s profile and the recommended items.
ILD	Link	It represents the diversity between the list of items recommended. This diversity is measured by the Pearson correlation of the item’s features vector.
Gini Coefficient	Link	Diversity is represented as the Gini coefficient – a measure of distributional inequality. It is measured as the inverse of cumulative frequency that each item is recommended.
Users Coverage	Link	It represents the percentage of distinct users that are interested in at least k items recommended (k ≥ 1).

API

For writing anything new to the library (e.g., value function, agent, etc) read the documentation.

Contributing

All contributions are welcome! Just open a pull request.

Related Projects

• iRec-cmdline