Blending absolute and relative time in autoencoders of variably-spaced sequences

Organisatsiooni nimi
TransferWise Ltd
Kokkuvõte
Much of real-life data, such as customer transactions at online retailers, can be represented as a series of transactions with uneven time periods between them. Predicting timing of future events is then key to measuring the total expected revenue from a particular customer. As long as we only consider time between events, without reference to calendar time, that can be easily modeled by an exponential or Weibull distribution, parametrized by an RNN fed by past events. Things get more interesting when we also want to capture the effect of calendar time, such as day of the week, on event likelihood.
This thesis will start with a purely relative-time neural network, for example as implemented in https://github.com/transferwise/neural-lifetimes, and add an overlay that allows fitting a calendar time dependence in a computationally-effective manner, to see if we can get the model to learn and reproduce both relative event spacings and calendar time-related behavior.
Lõputöö kaitsmise aasta
2022-2023
Juhendaja
Egor Kraev (Wise), Prof. Meelis Kull
Suhtlemiskeel(ed)
Nõuded kandideerijale
Tase
Magister
Märksõnad

Kandideerimise kontakt

 
Nimi
Egor Kraev
Tel
E-mail
egor.kraev@wise.com