Any data science and/or remote sensing experts have any insight pertaining to multivariate regression and deep neural networking with large feature set inputs and variable temporal data holes necessary for my dissertation work regarding climate change in the Arctic? I would greatly appreciate the feedback!
As an aside, I have utilized feature and stepwise extraction, most NaN mitigation methodologies, and VIF to mitigate multicollinearity prior to training. I’m simply seeking a deep learning solution that integrates all possible data to generate the best model.
Related to your query I suggest you to follow: https://www.mathworks.com/products/deep-learning.htmlArticle Temporal Convolutional Neural Network for the Classification...
https://www.nature.com/articles/s41598-018-24271-9As for as methodology of a research is concern it is almost similar, however., the differences may be based on data analysis, e.g., quantitative or qualitative etc.
For a PhD or a general scientific research project we may have more or less the following steps:
1) It may be unsolved (or partially solved, but may need further upgrades or improvements…may be open from decades to solve gradually, e.g., Human GENOME project etc., … ) topic or problem from academia, industry or other end users.
2) Crucial to know about the scope and gape of the research topic with getting questions on hands and the answers to be investigated.
3) In particular, it is important for the researcher/investigator to know about the importance and benefits of the underline research with goals to convince the academia, industry or other end users for getting research responsibility and grants.
4) To make literature review by reading and understanding most of the previous research findings related to the current research.
5) To evaluate the exact gape and then to gain further theoretical and computational knowledge with tools, related to the filling/solving of research gape or upgrade of the research questions at hand.
6) To outline the research steps and enhanced understanding and improvements in its development month by month within the given timeframe of the research accomplishment.
7) To accomplish and meet all the research goals (of the PhD/Project) and in some or many occasions to share it with the global scientific community by presentations to audiences and publications in conferences and journals.
8) The phase of research presenting to the live audiences and its publishing in conference and journals provides golden opportunities to get valuable discussions and comments from the other research experts.
9) By passing through the above phases, the PhD candidate or project solver has to write a complete report of their corresponding PhD research work (THESIS) or project report.
10) The report must have significant research impact that may be reviewed and acknowledged by some expert referees within the related research area.
Good Luck
Hello. You are working in an important area. It is my pleasure to suggest something.
I'm sure you're aware that the main benefit of deep learning approaches (compared with conventional machine learning) is that they do not require prior feature definition. On the other hand, they do require significant computational resources (GPU, lots of RAM, etc) and typically take only images as input. Thus, when you say you have a multivariate regression problem with temporal data gaps, you will first of all have to look at how the data can be "rasterized" into an image format. It is OK to use either grayscale or colour images, and most deep learning architectures can take at least a few sequential images as input - the easiest case being 3 successive images loaded onto the RGB channels of a colour image.
If you really want to try sequential techniques you need to look at RNN architectutes also, which typically take vector inputs but are designed to fit a sequence of data vectors. The GRU and LSTM approaches are the most successful ones since they allow for long process memory while avoiding numerical problems due to "vanishing gradients."
Another approach is GANs (generative adversarial networks), which are very good at learning multi-dimensional probability distributions for image data. In this type of approach, which has typically been used to fill gaps in images (inpainting), the deep neural network is trained to generate realistic images that a discriminator network cannot distinguish from the real data. Depending on the dimensionality of your data, you might be able to massage your problem into a format for which a GAN can be applied.
A simple initial approach might be to use segments of your data that do not have holes and develop a sequential predictor for this data. Then apply your predictor to the segments with holes to see how well you can bridge the gap.
Is there a "best" method? No. All deep learning and AI work is heuristic for now. There are not really any accepted design rules, although there are various benchmark data sets for image recognition. Performance metrics should be defined that are appropriate for your problem and keep your test data separate from your training/validation data. Data augmentation is usually helpful unless you have a lot of data and don't need to augment.
Please recommend if you find this answer useful.
- Badges
- Science topic