Just took the certificate exam but failed because I don’t know how to do multivariate time series forecasting. I’m quite disappointed that this topic is not covered in this course. Don’t want to complain too much. I just want to know where I can learn about this so I can get prepared next time.
2 Likes
Hi @Albert_Zhang,
are you talking of the last assignments of this specialization, or of the TensorFlow Developer Certificate Exam?
I’m saying that the last course does not cover multivariate time series forecasting which is required by the TensorFlow Developer Certificate Exam. Anyway, I just want to know where I can get this part of knowledge.
1 Like
I guess you’re talking of the 5th exercise in the TF Certification Exam. As far as I remember, it’s a time series question, and it should be mostly covered by the contents in the 4th course. Please consider that, even though this specialization covers a huge part of it, its goal is not to prepare for the TF Developer Certification. Anyway, I remember this course helped me a lot, even for the 5th exercise. I don’t know if they changed something in the last 7 months, since I’ve taken my certificate, but the handbook in the tensorflow website is still the same.
Best
I just finished this coursera coursera course and have run into the same problem as you. Did you end up learning multivariate time series forecasting? If so, how? What course did you take?
1 Like
In case anyone is interested, I learnt how to do one form of multivariate time series forecasting by:
-
attempting to read and understand this notebook: https://www.kaggle.com/code/jaimeggb/multi-variate-time-series-forecasting-tensorflow/edit. I failed, as it was too complex a level for me right now.
-
reading this: Time Series Forecasting as Supervised Learning. More specifically this part:
This was a simple explanation on how to do multivariate forecasting that was directly related to what I learnt for univariate forecasting on TF -
then grabbing the dataset from (1) and creating a notebook from scratch using the model and data processing methods I learnt in course 4 to do the Melbourne temperature assignment. This seemed simple at first, with minor changes, but I had to introduce and tweak several things (like scaling, and the model lambda layer for example), but when I managed to get a good forecast it felt like a great accomplishment.
2 Likes
Essentially, multivariate time series should be the very same, just with a different input shape. As specified in the lecture at corresponding steps (e.g., when defining a model and its first layer, input_shape is no longer (None, 1), but (None, NumberofFeatures), where features is another word for co-variates).
Additional concepts and techniques may become insightful for multivariate, eg feature selection (ex-ante analysis) and/ or feature importance (ex-post analysis). Also, different model architectures are better suited to deal with multivariate datasets than others. But, the general principles learned in the course, stay the same.
I hope this helps.
Cheers,
Steffen
–
All assertions are as of the best of my knowledge, no guarantee for correctness or completeness.
1 Like