SAP Data Scientist Interview Questions

Anonymous Hornet - "To handle the non-uniform sampling, I'd first clean and divide the dataset into chunks of n second interval 'uniform' trajectory data(e.g. 5s or 10s trajectories). This gives us a cleaner trajectory data chunks, T, of format (ship_ID, x, y, z, timestamp) to be formed. For the system itself, I'd use a generative model, e.g. Variational AutoEncoder (VAE), and train the model's 'encoder' to produce a latent-space representation of input features (x,y,z,timestamp) from T, and it's 'decoder' to pred"See full answer

Nick S. - "[I'm not sure whether the answer below is the best, as I have not gotten result and feedback from my interview] Ans: I would solve by first using a VAE-style model, to create a latent space embedding that translates user description to generate images. Training would be done on the 1000 avatar images and 100000 descriptions, following this scheme: VAE: description -> encoder -> latent space -> decoder -> image Q: "OK, but that means you're limiting the generated images to be only the 1000 imag"See full answer

Anonymous Hornet - "label #shape = [batchsize, posids] X #shape = [batch_size, ch] hidden = nn.Linear(batchsize, ch) #shape = [batchsize, ch] logits = nn.Linear(ch, C) #shape = [ch, C] #modifications start here neg_ids = torch.randint(0, C, 0.1 * C) #e.g. if we want to randomly sample only 10% of C outputneg = np.dot(logits[:, negids], hidden(X)) #shape [batch_size, C1] outputpos = np.dot(logits[:, posids], hidden(X)) #shape [batch_size, C2] output = torch.concat(outputpos, outputneg) #concatenate out"See full answer