Questions d'entretiens - Data analyst

un hipopotam

En Homme car au final c’est un animal comme les autres :) Et le seul aujourd’hui qui peut être data scientist !

Pas de réponse fixe, c'est juste pour mieux connaitre la personnalité du candidat.

Bonjour, Si possible, est-ce que vous pourriez me donner quelques questions techniques posées par Datalyo pendant votre entretien technique, s'il vous plait? En fait, j'aurais un rdv pour tester les compétences, mais je ne sais pas à préparer quoi? Merci Cordialement

J'ai répondu honnêtement (après un bac+5 en maths/stat je pense pouvoir m'orienter vers la data science)

I think estimating the probability P that a driver of Heetch would like to serve between two given locations could be one major problem. Eventually, Heetch would like to allocate drivers between two locations, based on this probability estimate. We can model p as a function of relevant factors: distance between points p, q, d=d(p,q), average income m_i of a driver, the corresponding standard deviation s_i of the incomes of the driver serving between p and q. We denote Heetch's revenue generated by these two points as R=R(p,q). Then, we can formulate the problem where we want to have a lower bound on R, but want to maximize the probability P of finding a driver willing to serve between p and q. To construct a suitable function P to maximize, we can study the covariance or correlation between each factor involved, and then form suitable f using this correlation coefficients. Next, we can maximize P, as a function of d(p,q), R, m_i, s_i, and solve the corresponding primal/dual problem as a constrained optimization problem on R>=R_min, for example.

Inhabituelle et posée en début d'entretien, la question peut déstabiliser mais il faut s'efforcer de ne parler QUE de ses loisirs, le recruteur (un analyste) fonctionnant beaucoup au feeling.

1. How do you use NN to reduce dimensionality? Approximate Nearest Neighbors using random projections is a dimensionality reduction technique. 2. Can you model time series as a linear regression model? Yes, using an autoregressive modeling technique. 3. a) Can you use resampling methods like bagging to estimate the max of a population? b) Why is bagging a variance reduction scheme? Bagging reduces variance b/c it trains many deep trees which when averaged together lowers variance while bias stays constant. 4. Why is the use of minibatch to minimize a function computationally more efficient than any other methods? Mini batch is a nice middle ground between an accurate gradient (batch GD) and memory constraints (SGD). Mini batches allow you to exploit data parallelism and also accurate gradient updates. 5.Gambler's ruin problem. Don't know. Maybe gamblers fallacy? 6. Assume that in a time series, some data are missing. How do you handle that? A. average out the existing values. Okay, so you want to average out the existing values, but how do you define the the new time series as a single function? A. Use characteristic or indicator function.