引導式測驗 · 硬 等級
XGBoost 演算法測驗
查看 XGBoost 的加性樹增強、客觀和正則化選擇、調整互動、缺失資料和公平評估。
問題 1 的 8
梯度提升通常如何建立其預測器?
預覽每個問題8 問題
當您參加測驗時,就會出現答案和解釋。
- 梯度提升通常如何建立其預測器?
- By averaging independent trees with unchanged fitting targets
- By replacing earlier trees entirely at every round
- By adding learners sequentially to improve the current objective
- By optimizing a single full-depth tree without additive updates
- 哪一對參數能捕捉每次加法更新大小與更新次數之間的折衷?
- Maximum depth and minimum child weight
- Learning rate and number of boosting rounds
- Column subsampling and row subsampling
- Leaf regularization and split-gain threshold
- 複雜度懲罰與樹限制扮演什麼角色?
- They guarantee probabilities are calibrated on every population
- They determine the ground-truth labels before fitting
- They replace the need to measure held-out performance
- They constrain model complexity and can reduce overfitting
- 為什麼早停需要單獨的最終測試評估?
- The monitored validation metric participates in choosing an iteration
- Early stopping makes the validation set unused during selection
- Stopping rounds should be chosen directly from final-test outcomes
- A chosen stopping point guarantees performance on any later population
- 關於 XGBoost 在表格資料上的哪項說法是合理的?
- It always beats every baseline
- It requires no data split
- It can be a strong candidate, but performance depends on the task and evaluation
- It guarantees production accuracy from a leaderboard result
- Why record the library version and objective with a trained model?
- The algorithm name alone fixes all defaults across releases
- Behavior and available options can vary by version and task setup
- An objective replaces the need to record the data split
- Native missing-value support establishes what every missing value means
- When are row or feature subsampling settings most defensibly chosen?
- By assuming any subsampling improves accuracy
- By selecting against final test labels
- By setting them without considering data size
- By measuring them within the validation protocol
- At training and inference time, what should remain consistent about the model inputs?
- Keep feature and missing-value conventions consistent with training
- Assign new category codes independently after deployment
- Recompute learned feature transformations separately for each inference batch without validation
- Change the meaning of missing values while retaining the same model artifact