Address binary time series prediction with highly imbalance dataset

[Jieyi1114]

I'm using DeepAREstimator to address a multiple binary time series problem. My target variable is highly imbalance (30:1). I set the threshold manually transforming the outputs to [0,1], but the performance is unsatisfactory and unstable.Can I modify the loss function to change the weights?

My current code is:
`

mx.random.seed(1)
np.random.seed(1)
prediction_length = 9
estimator = DeepAREstimator(
freq='W',
prediction_length = prediction_length,
context_length = 8,
use_feat_dynamic_real=True,
use_feat_static_real=True,
use_feat_static_cat=True,
cardinality=[26, 6, 2, 21], # Adjust cardinality values as needed
cell_type='lstm',
num_layers = 5,
num_cells = 40,
trainer = Trainer(epochs = 30, learning_rate = 0.01))
predictor = estimator.train(train_ds)
forecast_it, ts_it = make_evaluation_predictions(dataset=test_ds, predictor=predictor, num_samples=10)
...

`