Case study, part 2: tasks 9 to 14, hand-in, and the debrief

Step 1: Task 9: the Python node, because the subject says so15 min

The subject asks for a computed variable "using a Python Script widget". It is the only task in the whole exam that imposes a tool, and the mission checks it with transform_applied {op: "python"}.

1. Family Python, card Python (IA) (python).
2. The variable to build: how many items were recorded the same day in the same station. Reference: from 1 to 96, mean 11.1.
3. Write the function. The sandbox gives you pd and np, no imports and no file access:

def transform(df):
    # nombre d'objets enregistres le meme jour dans la meme gare
    cle = ["Date_year", "Date_month", "Date_day", "code_uic"]
    df["objets_meme_jour"] = df.groupby(cle)["Date"].transform("size")
    return df

The Describe the transformation field asks a language model to write the code for you. It needs a Gemini key on the instance; without one the panel says so and you write the four lines yourself. Either way you must be able to read them.

Step 2: Task 10: importance, and one explanation15 min

1. Open your best classifier, Explanations tab.
2. Permutation importance: about 4 seconds on a 2 000-row sample, baseline accuracy 0.687 for reference.
3. Partial dependence on the two strongest variables, about 0.4 s each.
4. Then the Predict tab, Simulator: load one item, move one variable, and use Quel changement fait basculer ?.

Write, in the Modélisation section, 500 characters minimum: which variables the model relies on, in which direction, and one concrete prediction explained end to end. No algorithm name, no jargon. The mission checks the length; your instructor checks the rest.

Step 3: Task 11: compare two models, on the right metrics15 min

1. Train a LightGBM (lgbm_clf) on the enriched dataset. Reference: accuracy 0.701, F1 0.620, AUC 0.764, about 8 seconds.
2. Train a Régression logistique (logreg). Reference: 0.684 accuracy, AUC 0.740, about 4 seconds.
3. Open the Comparison panel from the workshop header and read the two rows side by side.
4. Pour the comparison block into the Modélisation section.

Step 4: Task 12: the quality report as a deliverable12 min

1. On your enriched dataset: Explore → Quality, target returned. About 0.2 s on 258 809 rows and 21 columns.
2. Read and record: 2 924 duplicate rows, one constant column, the outlier columns, and the class balance.
3. Find the station that did not match the join: Boulogne Ville, 638 items. One station out of 156, and the reason is in the station file, not in your chain.
4. Pin the profile (Statistics tab, Pin) and pour it into the Données section.
5. Write 600 characters: missing values by column, duplicates, the unmatched station, and what you did about each.

And the reverse limitation: the Quality tab does not detect leaks. Leak detection happens at training time, on the pair (dataset, target). A student who only explores will never see the warning. That is a design choice, it is coherent, and you need to know it.

Step 5: Task 13: ethics and limits, with something in them13 min

Write the Conclusion et limites section, 700 characters minimum. Answer these, specifically, about this model:

1. Who would use this prediction, and to decide what? (the subject says: to organise storage and communication)
2. What happens to an item the model says will not be returned? Does the prediction become self-fulfilling?
3. Which variables would you refuse to use even if they improved the score, and why? (station, department, region are proxies for something)
4. The 2020 collapse is in your training data. What does that do to a model deployed in 2027?
5. What is the cost of a false positive against a false negative here, and does your metric reflect it?
6. What would you monitor after deployment?

Step 6: Task 14: the bonus that is worth the most15 min

Two enrichments, and the second one is the one that changes the exercise.

1. Time bands. Catégoriser (seuils) (categorize) on the hour, thresholds and labels of your choosing. One transformation.
2. Temporal split with a sealed test set. Filtrer des lignes (filter_rows) to build a 2019 to 2021 training set and a 2022 test set. Sceller the 2022 one. Then train with that sealed dataset as Jeu de test.
3. Compare with your random-split score. This is Session 8, on a real exam.

Step 7: Hand in, and the debrief20 min

1. Fill any of the seven sections still empty. The mission checks all of them by length; read mode shows you the gaps.
2. Pour the last blocks: metrics, confusion matrix, importance, the Démarche (lignage), and the Journal.
3. Reference for a complete answer: 7 sections out of 7, about 14 500 useful characters, 15 blocks, and an automatic journal of 87 entries.
4. Lecture, then Imprimer / PDF: the subject asks for a 3 to 6 page PDF organised by task number.
5. Vérifier one last time: 14 / 14, mission complete.

Debrief, together. The instructor puts the progress page on screen: one row per student, one column per step, hints consumed, last activity. Then discuss, with the numbers in front of you:

Question	What the trial found
Can this exam be taken in rakoon-ds without Orange?	Yes. Nine tasks out of ten are click-only; one task requires Python because the subject demands it
What is better than Orange here?	The report is inside the tool; the lineage and journal are automatic; leaks are flagged without being asked; the baseline is an algorithm, not a discipline; re-running nine transformations after re-plugging the head of the chain takes 0.05 s
What is worse?	Two forest hyperparameters missing; no cumulative distribution; no date difference in days; the graph becomes unreadable past fifty nodes and an exam makes seventy; a deep forest weighs 500 MB
What does the tool give away?	The weekday is a checkbox, the Coach already diagnoses imbalance and overfitting, and the Quality tab answers three tasks. The marking scheme has to move from 'could you build it' to 'which one did you use, and why'