TIME SERIES CLASSIFICATION | I Think I Failed This One 😔

This video explores a classification model, namely Multivariate LSTM-FCNs, which combines Long Short-Term Memory (LSTM) with a Convolutional Network (Fully Convolution Network or FCN). Then, we try to use this model to predict whether a company goes bankrupt or not. I'd like to share some other things I've tried but didn't get to record it: 1. Normalization. In neural networks, the common practice is to normalize the data before throwing it into the model. It's said over and over that normalization would make a considerable improvement on the model performance. Even though in this video I didn't do that (we only normalize the data through the BatchNormalization layer), I also have tried it, and the AUCROC score couldn't reach 0.5. 2. Use Precision-Recall Curve instead of AUCROC. AUCROC could be misleading if we have severely imbalanced data, so perhaps I should have showed you the precision and recall score, even though AUC under ROC curve could still be used as a metric. I tried using PrecisionRecallDisplay from sklearn and got the average precision score to be near 0, both from the weighted and unweighted model (with and without normalization). Now that I've finished editing this, a new thought come across my mind. If all modifications of the model still can't improve the metrics score, *then perhaps what we need is more data*. This might be one of those projects where we actually don't have enough data. I'd really like to hear what you think. Thank you so much for watching! Source code: https://www.kaggle.com/code/leesstephanie/time-series-classification-for-business 00:00 Intro 01:10 The model overview 02:57 LSTM & FCN 08:32 Squeeze and excite blocks 13:17 More details on the model Application with Python: 17:04 The data 21:35 The model (without class weights) 30:45 Weighted model 34:24 Do we have the winner? #timeseries #classification #timeseriesclassification #LSTM #convolution #machinelearning #pythonprogramming