Noise Injection

[michaeloc]

Firstly, congratulation on your work that is very interesting. However, I have some doubts. First, which is the Tf version that u use? Second, you generate two types of anomalies according to the paper (i.e, detour and route-switching), however going through your code I only find otype == 'random'(line 54 DataGenerator class) and shift (line 57) types. Are those types equal to used in the paper (random == detour and shift == route-switching)?

[liuyiding1993]

1. The tf version is 1.14.0.
2. Yes. random == detour and shift == route-switching

[michaeloc]

So, I have another question after going through your code: the results on evaluation are different from the paper even when I change eval_data to eval in run_evaluate. I have doubt wrt the variables, for example, is level equal to distance as you describe in the paper? Because the offset default walk one unit on the grid ([0,1],[1.0],[1,1]...etc) considering x and y, but you multiply by the level...so, if the level is 2, the final distance is like 200m since the geographical space is into 100mx100m? If I want to d=5, which is the variable to change (level=5)?

[michaeloc]

I also tried to verify the anomalies and compare them with the original trajectories, ex: plt.plot(anomalies[:,0],anomalies[:,1],c='red') # trajectory with anomalies generated by inject_outliers plt.plot(original[:,0],original[:,1],c='blue') # the original trajectory i.e, the trajectory above without anomalies

Overall, the shape of anomaly trajectories is very different from the paper pictures, this is correct, or did I execute something wrong?

[liuyiding1993]

So, I have another question after going through your code: the results on evaluation are different from the paper even when I change _evaldata to eval in run_evaluate. I have doubt wrt the variables, for example, is level equal to distance as you describe in the paper? Because the offset default walk one unit on the grid ([0,1],[1.0],[1,1]...etc) considering x and y, but you multiply by the level...so, if the level is 2, the final distance is like 200m since the geographical space is into 100mx100m? If I want to d=5, which is the variable to change (level=5)?

Yes, the level is to control the distance.

[liuyiding1993]

I also tried to verify the anomalies and compare them with the original trajectories, ex: plt.plot(anomalies[:,0],anomalies[:,1],c='red') # trajectory with anomalies generated by _injectoutliers plt.plot(original[:,0],original[:,1],c='blue') # the original trajectory i.e, the trajectory above without anomalies

Overall, the shape of anomaly trajectories is very different from the paper pictures, this is correct, or did I execute something wrong?

The paper pictures are all real data, not synthetic anomalies。

[michaeloc]

Hi, sorry by disturbing you. But I have some difficult to understand just one part of the code. The picture below show exactly the line. My question is: why is score_vals composed by two vectors (128 dimensions each). I know that 128 is the batch value then if I have 128 trajectories, 128 values should be returned, right? But why is the score returning 256 values?

[BruceLuke]

Hello, Michaeloc, Sorry to bother you, I just wanna know that could you give me your code about visulization! Thank you so much and hope you have a good day