ICACT20210141 Question.1
Questioner: kobiljonikromjanov@gmail.com 2021-02-09 오후 8:00:22 |
ICACT20210141 Answer.1
Answer by Auhor qiaobaiyou@mail.neu.edu.cn 2021-02-09 오후 8:00:22
Chrome  Click!! |
| Thank you very much for submitting your research article and welcome to IEEE/ ICACT2021. Can you please explain your contribution in this article? |
Our contribution can be summarized as follows:
1.We regarded SST field prediction as a spatiotemporal sequence forecasting problem and solved it using a CNN3D-LSTM model for the first time.
2. Based on the seasonal characteristics of the SST data, we proposed an ensemble learning prediction method, which use machine learning models to learn seasonal features, spatiotemporal model to learn the spatiotemporal features.
3. For the long-term SST prediction task, we introduce a custom attention layer behind the LSTM to adjust the finally lstm output and obtain better prediction results.
4. We have evaluated the proposed learning method on SST data and demonstrated its effectiveness. |
ICACT20210141 Question.10
Questioner: ch.lee@etri.re.kr 2021-02-10 오후 5:19:00 |
ICACT20210141 Answer.10
Answer by Auhor qiaobaiyou@mail.neu.edu.cn 2021-02-10 오후 5:19:00
Chrome Click!! |
| Thank you for your presentation. Did you conduct experiments with traditional methods for predicting time-series data? I am wondering how much your method is improved compared to those. |
Thank you for your question, traditional machine methods have been tried, but the results are not very good, because currently our prediction data is for a region of the data, for traditional machine learning methods, it is not easy to learn this spatial correlation. |
ICACT20210141 Question.2
Questioner: h.warunyu@gmail.com 2021-02-09 오후 12:56:18 |
ICACT20210141 Answer.2
Answer by Auhor qiaobaiyou@mail.neu.edu.cn 2021-02-09 오후 12:56:18
Chrome Click!! |
| Your research is very interesting. Now, I'm interesting in AI research. May I ask how long did you take to train the data with you equipment. |
Thank you for your question. This will depend on the size of the dataset. For the dataset used in the experiments, training the CNN3D-LSTM model for a long-term prediction task takes about 30 minutes or so under such equipment conditions, while training the ConvLSTM model can be done in about an hour. |
ICACT20210141 Question.3
Questioner: patcharin.ka@kmitl.ac.th 2021-02-09 오후 1:31:58 |
ICACT20210141 Answer.3
Answer by Auhor qiaobaiyou@mail.neu.edu.cn 2021-02-09 오후 1:31:58
Chrome Click!! |
| do you think is it has the limitation to apply you method in the real world? |
In the real world, if the method in the paper is used, it is better to have sufficient data and to conform to the periodic nature of time series in general. The method may be sensitive to the input of very volatile time-series data, and if the volatility is large, it is better to process it first by differencing, etc. In addition, the number of parameters in the model is quite large, and if the data set is small, overfitting may occur. |
ICACT20210141 Question.4
Questioner: h.warunyu@gmail.com 2021-02-09 오후 1:18:04 |
ICACT20210141 Answer.4
Answer by Auhor qiaobaiyou@mail.neu.edu.cn 2021-02-09 오후 1:18:04
Chrome Click!! |
| How can I use your method in real life? |
At First, you need to have a time-series prediction task like SST prediction or others so that you can get a reasonable dataset. The dataset is a structured dataset, in the case of sea surface temperature field prediction, the original dataset is m rows and n columns, each column represents an observation point in the field, each row represents a time point, and each value in it represents the temperature measurement at a point in time. With a sample of data, it is possible to extract temporal features based on the time of each row, and with the temporal features and temperature values, it is possible to extract the periodic patterns of the temporal data through the XGBOOST model, adding these as periodic features to the original data, and then, by cutting the data into 3D data through a sliding window method, the correct input to the model is obtained. Then, we can build the 3DCNN-LSTM model to complete the prediction task.
In fact, the method in the paper can be used not only for the task of sea surface temperature prediction, but also for other tasks of spatial and temporal series, such as rainfall prediction, video prediction, etc. |
ICACT20210141 Question.5
Questioner: tanatthep001@gmail.com 2021-02-09 오후 4:00:35 |
ICACT20210141 Answer.5
Answer by Auhor qiaobaiyou@mail.neu.edu.cn 2021-02-09 오후 4:00:35
Chrome Click!! |
| May I ask, how long did it take to train the data with your equipment? |
It takes about 30 minutes to train the CNN3D-LSTM model for a long-term prediction task and one hour to train the ConvLSTM model for the dataset used in experiments. |
ICACT20210141 Question.6
Questioner: alihussain.dream@gmail.com 2021-02-09 오후 4:13:56 |
ICACT20210141 Answer.6
Answer by Auhor qiaobaiyou@mail.neu.edu.cn 2021-02-09 오후 4:13:56
Chrome Click!! |
| Your research work is interesting, I like it. You have used real data or you have downloaded it from any site. |
You can download the real data from the Website https://psl.noaa.gov/data/gridded/. Thanks |
ICACT20210141 Question.7
Questioner: satyabrataaich@gmail.com 2021-02-09 오후 6:20:46 |
ICACT20210141 Answer.7
Answer by Auhor qiaobaiyou@mail.neu.edu.cn 2021-02-09 오후 6:20:46
Chrome Click!! |
| It's an interesting research work. How is your model perform compared to the previous state of art models? |
In the aspect of spatiotemporal sequence learning, ConvLSTM is a state of art multi observation point prediction model. Our model 3DCNN-LSTM-AT is superior to ConvLSTM model in terms of training efficiency and prediction accuracy. The corresponding experimental results are provided in this paper. |
ICACT20210141 Question.8
Questioner: quandh@vnu.edu.vn 2021-02-09 오후 8:04:22 |
ICACT20210141 Answer.8
Answer by Auhor qiaobaiyou@mail.neu.edu.cn 2021-02-09 오후 8:04:22
Chrome Click!! |
| Thank you for your great work. I have a question that what the differences between 3DCNN-LSTM-AT and 3DCNN-LSTM-AT-ONLY are (their results were mentioned in part 4). |
The two models are the same, but the training data are different. The former uses the time period features extracted by XGBOOST and the raw SST data, while the latter uses only the raw SST data. |
ICACT20210141 Question.9
Questioner: skdiakffn@ajou.ac.kr 2021-02-10 오후 5:33:33 |
ICACT20210141 Answer.9
Answer by Auhor qiaobaiyou@mail.neu.edu.cn 2021-02-10 오후 5:33:33
Chrome Click!! |
| In this paper, the 3DCNN-LSTM-AT technique shows the highest accuracy. However, looking at the r^2 indicator alone, it seems that there is no significant difference in accuracy from other techniques. What is the reason for this? In addition, there is no significant accuracy difference, especially with the 3DCNN-LSTM method, but the learning time seems to be close to three times. Therefore, is it better to use the 3DCNN-LSTM method when there is a lack of computation or learning time? |
The r^2 is not really a measure of the model's prediction accuracy, but rather it is something like describing how confident the model is in predicting the final outcome, and the model with the greater confidence and accuracy is often the best. Therefore, it is often used as a reference rather than a rigid indicator to evaluate the performance of a model. If the model predicts a result that differs significantly from the true result but the model is still confident, r^2 is often high. For 3DCNN-LSTM-AT, if it is a long-term prediction task, the result is often better than that of 3DCNN-LSTM, because the Attention mechanism plays a bigger role, after all, the historical input sequence will be longer, while if it is a common short- to medium-term prediction task, the improvement effect may not be so obvious. |
