น้ำหนักคลาสสำหรับข้อมูลไม่สมดุล

If 95 percent of your samples are one class, the model can score high by always guessing it. That is the class imbalance trap. ⚖️

A lazy model that ignores the rare class still looks 95 percent accurate. So accuracy hides total failure on the minority you actually care about.

Each sample contributes equally to the loss, so the abundant class pulls the gradient hardest. The model learns mostly from the majority class.

The fix is to give rare classes a heavier voice. Class weights multiply each sample's loss so minority mistakes count for more.

A common recipe sets each weight inversely proportional to its class frequency. The rarer the class, the larger its weight grows.

For multiclass tasks, hand a weight tensor of length C to the loss. Each entry scales the penalty for that class's errors.

For two-class problems, BCEWithLogitsLoss takes pos_weight, a single number that boosts the positive class to offset its rarity.

Weighting is not the only tool. You can also oversample the rare class or undersample the common one to balance each batch.

A WeightedRandomSampler draws minority samples more often, so every batch the DataLoader serves is roughly balanced by design.

For extreme imbalance, focal loss down-weights easy, well-classified examples so the model focuses on the hard, rare ones.

After rebalancing, drop accuracy and track F1 or precision and recall. They reveal whether the rare class is finally being learned.

Your dataset is 95 percent negative. How do you stop the model from ignoring positives?

Imbalance lets the majority hijack training, so you fight back with class weights, balanced sampling, or focal loss, and you judge results with F1. 💪