第 4 / 6 页
非代码面试题
显示 20 / 108 道匹配题目
答题状态：未尝试未正确已正确
ID题目领域难度题型进度权限
2532Hessian of Logistic Negative Log-Likelihood 4For one observation with score z = w T x, what is the Hessian of the negative log-likelihood with respect to w?机器学习简单derivation未尝试免费2534One Gradient Step on a Tiny Logistic ProblemA one-feature logistic model without intercept uses beta = 0 initially, learning rate 0.2, data x = [-1, 0, 1], and labels y = [0, 0, 1]. What is beta after one gradient step on the negative log-likelihood?机器学习困难数值题未尝试面试订阅2535Decision Threshold Under Asymmetric Classification CostsA desk incurs cost 1 for a false positive and cost 5 for a false negative. Under a calibrated logistic probability p = P(Y=1|x), above what threshold should it predict class 1 to minimize expected cost?机器学习困难derivation未尝试面试订阅2536Why Softmax Fixes Joint Normalization 16What does the softmax construction add that one-vs-rest logistic models do not provide automatically?机器学习简单essay未尝试免费2537Why Logistic Probabilities Are Useful Downstream 18Why is it valuable that logistic regression produces a calibrated probability estimate rather than only a hard class label?机器学习中等essay未尝试面试订阅2538Why Logistic Beats Hard Threshold Rules for Training 23Why is a smooth probabilistic loss easier to optimize than training directly against a hard classification rule?机器学习中等essay未尝试免费2540Intercept Shift for a Deployment Prior ChangeA logistic model was trained under class prior 0.5 and has intercept -0.4. At deployment the base rate falls to 0.2 while feature likelihood ratios are assumed unchanged. What adjusted intercept should be used?机器学习困难数值题未尝试面试订阅2541One Gradient Step on a Single Logistic Observation 22For one observation with x = 2, y = 1, current weight w = 0, and learning rate eta = 0.4, what is one gradient-descent update on the negative log-likelihood?机器学习简单数值题未尝试免费4166Centering-and-Scaling Coefficient Rewrite 1A linear model is y = 1.5 + 2 x. You now replace x by the engineered feature z=(x-10)/2. What intercept and slope make the model equivalent when written as y = a + b z?机器学习简单数值题未尝试面试订阅4167Centering-and-Scaling Coefficient Rewrite 2A linear model is y = -0.5 + 1.2 x. You now replace x by the engineered feature z=(x-5)/0.5. What intercept and slope make the model equivalent when written as y = a + b z?机器学习简单数值题未尝试面试订阅4171Marginal Effect with an Interaction Feature 1A model uses engineered interaction terms: y = β0 + 0.8 x1 + β2 x2 + 0.5 x1 x2. What is the marginal effect of x1 when x2 = 2?机器学习中等数值题未尝试面试订阅4172Marginal Effect with an Interaction Feature 2A model uses engineered interaction terms: y = β0 + -0.2 x1 + β2 x2 + 1.2 x1 x2. What is the marginal effect of x1 when x2 = -1?机器学习中等数值题未尝试面试订阅4176Cyclical Time-of-Day Encoding 1A cyclical hour-of-day feature is encoded as (sin(2πh/24), cos(2πh/24)). What is the encoding for h=6?机器学习简单数值题未尝试面试订阅4177One-Hot Plus Interaction Column CountA categorical variable has 5 levels. You one-hot encode it with a dropped baseline, keep one raw numeric feature x, and also create all interactions between x and the retained dummies. How many columns come out of this block in total?机器学习简单数值题未尝试面试订阅4178Winsorize-then-Standardize PipelineA raw daily return of 4.8% is winsorized to the range [-3%, 3%], then standardized using trailing mean 0.5% and trailing standard deviation 1.0%. What z-score feature results?机器学习简单数值题未尝试面试订阅4179Log1p Volume TransformA liquidity feature uses log1p(volume). If today's volume is 999999 shares, what transformed value do you store?机器学习简单数值题未尝试面试订阅4180Leakage-Safe Rolling Mean FeatureAt today's open you build a leakage-safe rolling-mean return feature from the last four completed daily returns: [1.0%, -2.0%, 0.5%, 1.5%]. What feature value do you use?机器学习简单数值题未尝试面试订阅4181Rolling Mean with a Forward WindowA candidate constructs a '5-day moving average' for day t using prices from t-2 through t+2. Is this valid for a predictive model meant to trade at the close of day t?机器学习中等derivation未尝试面试订阅4182Standardization Before the Train-Test SplitA pipeline standardizes each feature using the mean and standard deviation of the full dataset before creating the train/test split. Is that clean?机器学习中等derivation未尝试面试订阅4183Target Encoding Without Out-of-Fold LogicA categorical feature is target-encoded using the full sample average label for each category, and then those encodings are used inside cross-validation. Is that safe?机器学习中等derivation未尝试面试订阅