Ensemble Bandit

This repository contains the code and data needed to reproduce the experiments of the paper:

R. Cañamares, M. Redondo, P. Castells. Multi-Armed Recommender System Bandit Ensembles. 13th ACM Conference on Recommender Systems (RecSys 2019). Copenhagen, Denmark, September 2019.

The software produces the results displayed in figures 1, 2 and 3 in the paper.

Authors

Information Retrieval Group at Universidad Autónoma de Madrid

• Rocío Cañamares (rocio.cannamares@uam.es)
• Pablo Castells (pablo.castells@uam.es)

Software description

This repository contains all the needed classes to reproduce the experiments reported in the paper. The software contains the following packages:

• es.uam.ir.ensemblebandit.arm: classes implementing bandit arms.
• es.uam.ir.ensemblebandit.bandit: classes implementing different bandit strategies.
• es.uam.ir.ensemblebandit.ensemble: class implementing a dynamic ensemble.
• es.uam.ir.ensemblebandit.datagenerator: classes to handle the data and generate the different training sets for each algorithm.
• es.uam.ir.ensemblebandit.filler: classes to complete recommendation rankings when an algorithm falls short of coverage.
• es.uam.ir.ensemblebandit.util: additional classes, useful for the rest of the program.
• es.uam.ir.ensemblebandit: top-level main classes to generate the figures of the paper.

The software uses the RankSys library, and extends some of its classes. Our extensions are located in the following package:

• es.uam.ir.ensemblebandit.ranksys.rec.fast.basic: extension of RankSys implementations of non-personalized recommendation, adding popularity-based recommendation.

System Requirements

• Java JDK: 1.8 or above (the software was tested using the version 1.8.0_181).

• Maven: tested with version 3.6.0.

Installation

Download all the files and unzip them into any root folder.

From the root folder run the command:

mvn compile assembly::single

Execution

To run the experiments that produced the results displayed figures 1 and 2 of the paper, run the command:

java -cp .\target\EnsembleBandit-0.1-jar-with-dependencies.jar es.uam.ir.ensemblebandit.Figure1and2 dataPath

Where dataPath is the rating data, including one rating per line with the format: user \t item \t rating.

Three files will be generated into the root folder: figure1.txt, figure2-epsilon-greedy.txt and figure2-thompson-sampling.txt.

• figure1.txt contains the cumulated recall achieved by different recommender systems (ensembles and standalone algorithms) up to each epoch.
• figure2-epsilon-greedy.txt and figure2-thompson-sampling.txt contain the number of times each arm has been selected by the respective ensemble bandit (ε-greedy and Thompson sampling) at each epoch.

To generate the experiment for the results displayed in figure 3 run the command:

java -cp .\target\EnsembleBandit-0.1-jar-with-dependencies.jar es.uam.ir.ensemblebandit.Figure3 dataPath

Where dataPath is the rating data, including one rating per line with the format: user \t item \t rating.

A file figure3.txt will be generated inside the root folder, with the cumulated recall achieved by different recommender systems (ensembles and standalone algorithms) at each epoch.

Example of the output files

Exact values slightly change from one execution to another:

• figure1.txt

    Epoch	Random recommendation	Most popular	User-based kNN	Matrix factorization	Thompson sampling ensemble	Epsilon-greedy ensemble	Dynamic ensemble
    0	2.696704554150921E-4	0.004934240495027496	3.1704499487990556E-4	4.701011993046875E-4	0.0020225284156131906	0.0018548954298146197	4.701011993046875E-4
    1	5.959866586472744E-4	0.009523722327003718	7.746371052996114E-4	6.362111710660639E-4	0.006815675505926515	0.006465863453815261	6.963686743473822E-4
    2	8.769514905440563E-4	0.013965714832317606	0.0013621618073170153	8.91572723344437E-4	0.011313831787702835	0.010842215685267644	9.827947107775565E-4
    3	0.001121573047231903	0.017744897695232676	0.002100092871179099	0.0014261510881241007	0.015549588962334908	0.014987861618770101	0.001535714481130957
    4	0.001402765034120572	0.021789122685983674	0.0029178314559759853	0.0023794057324701295	0.019419667641289464	0.018824408477457134	0.002500109493846446
    5	0.0016987532136725002	0.025713642008754926	0.003859508058915975	0.0028620272723619855	0.02345613422399893	0.022774023466508243	0.0030158115473981092
    6	0.001963908616192755	0.02911732161014415	0.0048746492630408296	0.0035906741148028967	0.027161962808029228	0.02656038288997691	0.0036204668775711986
    7	0.002230977357953197	0.033234566049058584	0.006018460028600044	0.004246113672230014	0.030685778674563034	0.02989896495668457	0.004149316889676448
    8	0.0024616688215399673	0.03707239369153599	0.007231486386658137	0.00478686028808994	0.03480872387420552	0.03279251550667655	0.004784566434116612
    9	0.0027343964052824812	0.04087283063298364	0.008548781092107702	0.005393608957858082	0.03890360843697257	0.035747019356538595	0.005443031253947708
    10	0.002994586357914277	0.0451088884294307	0.009861070273448985	0.00601606237359674	0.042791581785273425	0.039347028447182464	0.0061830739495644365
    ...
  

• figure2-epsilon-greedy.txt and figure2-thompson-sampling.txt

    Epoch	Most popular	User-based kNN	Matrix factorization
    0	1893	1865	1987
    1	5637	194	202
    2	5630	219	189
    3	5596	215	228
    4	5605	221	211
    5	5622	205	204
    6	5692	169	177
    7	5644	205	188
    8	5654	188	192
    9	5608	219	209
    10	5655	184	197
    ...
  

• figure3.txt:

    Epoch	Random recommendation	Most popular	User-based kNN	Matrix factorization	Dynamic ensemble
    0	2.847825895878563E-4	0.004953810725052957	2.7950883792882196E-4	4.6057431155566885E-4	0.004953810725052957
    1	5.433511578303232E-4	0.008540077768767767	6.541280402886603E-4	6.032434447838788E-4	0.008540077768767767
    2	7.914926848487194E-4	0.013169722403953045	0.001153948315077144	8.372734436564224E-4	0.013169722403953045
    3	0.0010555853660682085	0.01660946124785591	0.0017775462470014907	0.0013125045084353927	0.01660946124785591
    4	0.0013268690992600154	0.02065840973674334	0.0024707446574329262	0.0022460478534929037	0.02065840973674334
    5	0.0016335123507967774	0.02459016393442623	0.003290146126897077	0.002744992327059317	0.02459016393442623
    6	0.0019439118055665605	0.028534578880340663	0.004252294581190431	0.003291045078680127	0.028534578880340663
    7	0.0022263280910117286	0.032349994756009796	0.005315618137009131	0.0038713341375831586	0.032349994756009796
    8	0.0024806594889630034	0.03698057045365858	0.0064081266214716425	0.004461228955634772	0.03698057045365858
    9	0.002768540498055325	0.04122301633605601	0.007577459755249642	0.005164860584079516	0.04122301633605601
    10	0.0030355166019136797	0.04547043908868731	0.008787187462360868	0.005832876857561889	0.04547043908868731
    ...