Pergunta de entrevista da empresa Amazon
Task: 300-d feat vector of sensor data come in every second -- decide whether or not there is a problem. The task was Anomaly Detection. They did not give me any other information but you were allowed and even encouraged to do so. The next Question concerned the type of problems. There are 5 type of problems but we need to identify if there is a new one. The last Question concerned production. What should be shared across the different servers?
Resposta da entrevista
Sigiloso
Q1: I chose to say Gaussian Mixture Model and talked about how to find the number of Gaussians. The next question concerned the fact that these metrics oftentimes can be a mistake -- a false alarm, what do you do then? I answered that I would collect 60 points (1' data) and then decide if N alarms are present within this timeframe then there is a problem. Of course the numbers 60 and N are parameters that I would decide only after I investigate the data they have been collecting for 2 years. At this point it was clear that the interview was going well. Q2: Novelty Detection. I started with GMM or Single SVM -- to which the interviewer responded that he would have tried this himself. Q3: the parameters of GMM; mean and covariance of each Gaussian that is. I struggled a lot since I do not have experience on these things. He gave away the answer at the end.