Stepan Balcar, Peter Vojtas and Michal Kopecky Charles University - - PowerPoint PPT Presentation

Stepan Balcar, Peter Vojtas and Michal Kopecky Charles University e-mail(s): <name>.<surname>@mff.cuni.cz

Outline

— Introduction — Nature-inspired algorithms for matrix factorization — Island models — Results — Future Work

DA2PL 2018 – Balcar, Vojtas , Kopecky Synergy of stochastic and … in recommender systems 2

Recommendation

The problem Illustration

— Recommendation of items to users

based on their preferences

— We have chosen domain of movies

— MovieLens datasets

— ML-100k 100000 ratings by 943 users on

1682 movies

— ML-1M 1000209 ratings, by 6040 users on

3900 movies

— Usable also (and not only) for

recommendation of

— Music, books, recipes, holidays

DA2PL 2018 – Balcar, Vojtas , Kopecky Synergy of stochastic and … in recommender systems 3

Databse of items and preferences

Input:

Recommended items

Result: Recommendation User + profile

Our approach

Our model Matrix factorization

— Collaborative filtering

— If possible, content-based filtering

— Based on matrix factorization

— Extraction of latent vectors

• f users and items

— Model:

— Pair of latent vectors

— Nature inspired computation

— Genetic algorithms — Neural networks?

DA2PL 2018 – Balcar, Vojtas , Kopecky Synergy of stochastic and … in recommender systems 4

Evolutionary algorithms and SGD

— Individual

= pair of latent vectors

— Operators based on SGD

— Example:

— Cross = One-point cross — Mutation = SGD

5

• train set → count SGD-gradients
• train set → evluation

DA2PL 2018 – Balcar, Vojtas , Kopecky Synergy of stochastic and … in recommender systems

Distributed computing

— More approaches

— Master-slave model — Cellular model — Island model

6 DA2PL 2018 – Balcar, Vojtas , Kopecky Synergy of stochastic and … in recommender systems

Co-evolution and Migration

— Individuals from one island can

time to time migrate to another island

— Mixing of “genetic material”

— Dynamic re/planning

— Island can change parameters, or even the used method

— Stochastic: Hill climbing, Random search, Simulated annealing, Tabu search, — Evolutionary: Brute force, Differential evolution

— Based on external planner

— Possibility to take advantage of different approaches to learning

to gain advantage

— Mutual help of methods — SYNERGY

7 DA2PL 2018 – Balcar, Vojtas , Kopecky Synergy of stochastic and … in recommender systems

Experiments

8 DA2PL 2018 – Balcar, Vojtas , Kopecky Synergy of stochastic and … in recommender systems

Results

— si-*

— Single method

— par-*

— Parallel approach

— coe-*

— Co-evolution of

more methods

DA2PL 2018 – Balcar, Vojtas , Kopecky Synergy of stochastic and … in recommender systems 9

Preliminary NN results

—

At this moment, NN approach seems to outperform our results

—

NN available (https://github.com/zishansami102/Recommendation- Engine) and their modifications achieve RMSE under 0,85 – better than 0,88

—

Our preliminary tests of the NN behavior show that around 20th epoch the best results were achieved.

—

Some settings achieved RMSE around 0.835.

—

Further training shows overfitting.

—

While the RMSE on training data continues its descent, the RMSE on testing data starts to grow back to the 0.85 level. —

Varying top-k (originally 5) considered ratings per user does not substantially change the picture.

—

The optimum is moved to higher epochs.

—

Looks promising,

—

The lost of generalization capability on ml-1m dataset might mean the ability to learn user preferences on bigger datasets

DA2PL 2018 – Balcar, Vojtas , Kopecky Synergy of stochastic and … in recommender systems 10

Future work

— Utilize the maximum potential of cooperation between different methods

— Ensemble — Meta/hyper heuristics

— Create a dynamic adaptive portfolio for online recommendations — Use of existing optimized library implementations of methods whenever

possible for speedup of computation

— Involving NN somehow in the island model

— Introduce “Deep School Island”, that

— will prepare promising individuals for other islands/algorithms — ? will adopt promising individuals from outside and try to optimize them further ?

— …

11

Previous related publications

— Concerned on recommenders

— Balcar Š.: Influence of the individual's size on the island model architecture, in Digital Library University of

West Bohemia, Brno, University of West Bohemia, ISBN: 978-80-214-5679-2, pp. 163-168, 2018

— Balcar Š.: Preference learning by matrix factorization on island models, in Proceedings of the 18th Conference

Information Technologies - Applications and Theory (ITAT 2018), Krompachy, CEUR Workshop Proceedings, ISBN: 978-1-72726-719-8, ISSN: 1613-0073, pp. 146-151, 2018

— Concerned on combinatorial optimization (TSP, ...)

— Balcar Š., Pilát M.: Heterogeneous Island Model with Re-planning of Methods, in GECCO 2018 Companion -

Proceedings of the 2018 Genetic and Evolutionary Computation Conference Companion, Kyoto, Association for Computing Machinery, Inc, ISBN: 978-1-4503-5764-7, pp. 245-246, 2018

— Balcar Š., Pilát M.: Online Parallel Portfolio Selection with Heterogeneous Island Model, in International

Conference on Tools with Artificial Intelligence, Volos, Greece, 2018

12

Questions?

ITAT 2017 - Kopecky, Vomlelova, Vojtas Repeatable Web Data Extraction and Interlinking 13

Implementation of island models: https://github.com/sbalcar/distributedea/

14