Service demand data (clean)
A service-demand prediction dataset: case_id, created_at, features_*, target_numeric, target_binary, group_attribute, split_time. Used in the Guided Lab.
DS 201 Practice Center
Use synthetic service-demand prediction data to practice framing a prediction problem, hunting for leakage, building baselines, choosing metrics, trading off thresholds, and reviewing error slices.
Info
The DS 201 practice data is synthetic teaching data; every site, case, and value is illustrative and does not represent any real organization or person.
Warning
Use only the synthetic files provided by the course. Unless authorized, never paste confidential institutional or company data, personal data, or other sensitive information into an AI tool; check your institution's policy before working with real operational data.
Download the practice files
Service demand data (practice, with challenges)
The same schema, with deliberate judgment challenges: one column that looks useful but is actually leakage, a baseline scenario that is easy to misuse, and a subgroup with a noticeably higher error pattern. Used in the Practice Center and the Independent Lab.
Data dictionary (Traditional Chinese)
Column types, units, grain, allowed values, and provenance for the service-demand data file (Traditional Chinese).
Data dictionary (English)
English data dictionary; columns and units match the Chinese version.
Guided Lab starter notebook
A starter notebook with instructions and items to complete, runnable top-to-bottom from a fresh kernel, producing a baseline vs. first-model comparison.
Independent Lab starter notebook
A starter notebook and checklist template for the Predictive Analytics Use Case + Model Card.
Practice tasks
Practice A: Prediction Framing Canvas
For the practice data, write down a clearly framed prediction problem: prediction unit, target, decision, and which columns are available at prediction time.
Files used: Service demand data (practice, with challenges), Data dictionary (Traditional Chinese)
Steps
- Choose target_numeric or target_binary as the prediction target and state the business decision it supports.
- List every features_* column and mark, one by one, whether it existed before split_time.
- Write down exactly what the prediction unit (case_id) represents: which site, which time window.
- State in one sentence what happens if this prediction is wrong.
What you produce
- A completed Prediction Framing Canvas
- A per-feature mark of availability at prediction time
- A one-sentence statement of the consequence of a wrong prediction
Practice B: Leakage Hunt
One column in the practice data looks highly predictive but is actually target leakage. Find it and explain why it cannot be used.
Files used: Service demand data (practice, with challenges), Data dictionary (Traditional Chinese)
Steps
- Check the data dictionary description of every features_* column and find one produced later than split_time.
- Use a correlation or a simple model to confirm the column's predictive power is suspiciously high.
- Write down why this column constitutes leakage, and how you expect model performance to change once it is removed.
- Propose at least one reusable checklist item for judging whether a column is safe to use in future projects.
What you produce
- The leaking column found and the reasoning behind it
- A before/after comparison of model performance
- A reusable leakage checklist
Practice C: Baseline Challenge
Build at least one baseline for each of the practice data's two targets (target_numeric and target_binary), and explain your choice.
Files used: Service demand data (practice, with challenges), Data dictionary (Traditional Chinese)
Steps
- Build a mean baseline for target_numeric and compute its MAE.
- Build a majority-class baseline for target_binary and compute its accuracy.
- Try at least one rule-based baseline (for example, the mean by group_attribute) and compare it to the mean baseline.
- Write a short statement of what minimum performance a new model must reach to be "worth using."
What you produce
- Baseline performance for both targets
- A comparison of the rule-based baseline against the mean baseline
- A written statement of the "worth using" threshold
Practice D: Metric Selection
For a classification scenario, state whether you would prioritize precision or recall, and back your choice with data.
Files used: Service demand data (practice, with challenges), Data dictionary (Traditional Chinese)
Steps
- Pick a business scenario (for example: missing a site that really will exceed capacity costs far more than a false alarm).
- Compute how different thresholds affect precision and recall in this scenario.
- Explain why this scenario should prioritize recall or precision.
- Write down the worst-case consequence of prioritizing the wrong metric.
What you produce
- A precision/recall table across thresholds
- A written rationale for the metric choice
- A worst-case analysis of choosing the wrong metric
Practice E: Threshold Trade-off
Build the confusion matrix at three or more thresholds and explain how the threshold changes false alarms and missed cases.
Files used: Service demand data (practice, with challenges), Data dictionary (Traditional Chinese)
Steps
- Compute the confusion matrix at three or more thresholds (for example 0.3, 0.5, 0.7).
- Compile a table comparing False Positives and False Negatives across thresholds.
- Recommend a threshold based on the business scenario defined in the Metric Selection task, with reasoning.
- State the conditions under which this recommended threshold should be revisited (for example, when the data distribution shifts).
What you produce
- Confusion matrices at three or more thresholds
- A recommended threshold with reasoning
- Trigger conditions for revisiting the threshold
Practice F: Error Slice Review
Check by group_attribute or demand range whether model error concentrates in a specific subgroup, and raise questions to investigate before deployment.
Files used: Service demand data (practice, with challenges), Data dictionary (Traditional Chinese)
Steps
- Train a simple model and compute the error for every case on validation.
- Group by group_attribute and compare whether average error is consistent across groups.
- Identify the subgroup or situation with the largest error and write down at least one hypothesis for why.
- List the concrete implications of this finding for the deployment decision (cost, monitoring, human fallback).
What you produce
- An error table broken down by subgroup
- A hypothesis for the highest-error subgroup
- Concrete recommendations for the deployment decision
The done mark is stored in this browser only (and synced once you join a cohort). Keep your work in your own files or notebook.