Novel Learning and Exploration-Exploitation Methods for Effective Recommender Systems

Romain Warlop 1
1 SEQUEL - Sequential Learning
Inria Lille - Nord Europe, CRIStAL - Centre de Recherche en Informatique, Signal et Automatique de Lille (CRIStAL) - UMR 9189
Abstract : This thesis, written in a company as a CIFRE thesis in the company fifty-five, studies recommender systems algorithms. We propose three new algorithms that improved over state-of-the-art solutions in terms of performance or matching industrial constraints. To that end, we proposed a first algorithm based on tensor factorization, a generalization of matrix factorization, commonly used on collaborative filtering. This extension allows to take into account simultaneously several types of feedbacks as well as different contexts. The proposed algorithm is also highly parallelisable thus suitable for real life large datasets. We then proposed a new algorithm that improves basket completion state-of-the-art algorithms. The goal of basket completion algorithms is to recommend a new product to a given user based on the products she is about to purchase in order to increase the user value. To that end we leverage Determinantal Point Processes (DPP), i.e., probability measure where the probability to observe a given set is proportional to the determinant of a kernel matrix. We generalized DPP approaches for basket completion using a tensorial point of view coupled with a logistic regression. Finally, we proposed a reinforcement learning algorithm that allows to alternate between several recommender systems algorithms. Indeed, using always the same algorithm may either bore the user for a while or reinforce her trust in the system. Thus, the algorithm performance is not stationary and depends on when and how much the algorithm has been used in the past. We then model the future performance of an algorithm according to linear function which is a polynomial in a recency function, that is a function that measures the frequency of use of an algorithm in a recent history. Our reinforcement learning algorithm learns in real time how to alternate between several recommender system algorithms in order to maximize long term performances, that is in order to keep the user interested in the system as long as possible. This algorithm can be seen as an hybrid recommender system. This thesis having been written in a company, we always looked for considering industrial contraints when developing new algorithms. Thus, each chapter that introduces a new algorithm will contain a section in which we present how the solution has been used or could be used in practice.
Document type :
Theses
Complete list of metadatas

Cited literature [141 references]  Display  Hide  Download

https://hal.inria.fr/tel-01915499
Contributor : Romain Warlop <>
Submitted on : Wednesday, November 7, 2018 - 4:30:58 PM
Last modification on : Friday, May 17, 2019 - 11:39:16 AM
Long-term archiving on : Friday, February 8, 2019 - 3:29:43 PM

File

Rapport de thèse - VF - Romai...
Files produced by the author(s)

Identifiers

  • HAL Id : tel-01915499, version 1

Citation

Romain Warlop. Novel Learning and Exploration-Exploitation Methods for Effective Recommender Systems. Artificial Intelligence [cs.AI]. Lille1, 2018. English. ⟨tel-01915499⟩

Share

Metrics

Record views

282

Files downloads

411