English:Recommendation systems and feedback loops - AI Year 6
Introduction

Why do the same kinds of suggestions keep appearing? Trace three recommendation rounds, change one signal and compare explanation, reset, browsing and a diversity cap. The catalogue and clicks are invented; no real learner profile or portal setting is changed.
Read signals and define a useful recommendation
A recommendation is an inference from selected signals
A reading app may suggest more mystery books after several mystery clicks. That does not reveal a person's complete or permanent preferences. A click might reflect curiosity, an assignment, an accidental tap or the options that happened to be visible. Content features describe items; user signals describe observed or explicitly supplied interactions. Both are incomplete evidence for what someone needs now.
Search begins with an explicit information need or query; recommendations may suggest items without a new query. A related-item panel is not necessarily the same as a personalised home page. Ask what the system is trying to rank: likely clicks, usefulness for a stated lesson, variety or another declared objective. High rank is not a certificate of accuracy, trustworthiness or age suitability.
Our experiment is a transparent scoring simulation with six invented book labels. They are original classroom examples, not a claim that these books exist or appear in the credited library photograph. No pupil reading history, real account or personal profile is collected. The program uses a prepared counter, not a trained commercial recommender.
Read the whole candidate catalogue
| ID | Invented title | Genre | Frozen base click count |
|---|---|---|---|
| B1 | The Lantern Clue | mystery | 8 |
| B2 | The Clocktower Note | mystery | 7 |
| B3 | The Missing Map | mystery | 6 |
| B4 | Letters from an Old Harbour | history | 3 |
| B5 | Small Experiments Big Questions | science | 2 |
| B6 | Rain on the Window | poetry | 1 |
The base counts are fictional, fixed popularity signals. Initially a fictional profile also has mystery_count two and zero for the other genres. Our personalised score adds the base count to twice that genre counter. B1 therefore scores twelve, B2 eleven and B3 ten; B4 scores three. Equal scores are ordered by item ID. The score is not a probability, reading level or source-quality rating.
CodeAI: AI: What is Machine Learning? With teacher support, compare patterns inferred from examples with our deliberately simple hand-written score. The wider publisher unit is for older grades. Our Year 6 simulation needs no live pupil AI account, real click collection or access to the photographed library's systems.
Keep the real user goal visible
If the current task is to find historical information, a repeated mystery click does not cancel that need. The learner can state a history goal, ask a teacher or browse the catalogue directly. A useful system should make the basis of a suggestion understandable and offer a way to change the situation. Our simulation makes those choices inspectable; it does not change any real portal settings.
Trace three feedback rounds and change exposure
Simulate three rounds of exposure and feedback
Popularity-only ranking uses the frozen base counts and gives B1, B2, B3. Personalised ranking initially gives the same three IDs with scores twelve, eleven and ten. For this controlled simulation only, the prepared user clicks the first personalised result, B1, each round. Increase the mystery profile counter by one after the click; do not change the frozen base counts.
| Round | Mystery counter before display | Personalised top three | Their scores | Counter after prepared click |
|---|---|---|---|---|
| 1 | 2 | B1 B2 B3 | 12 11 10 | 3 |
| 2 | 3 | B1 B2 B3 | 14 13 12 | 4 |
| 3 | 4 | B1 B2 B3 | 16 15 14 | 5 |
The next scores after round three would be eighteen, seventeen and sixteen. We have not observed real preferences: the click behaviour was supplied by the exercise. The loop is exposure, prepared click, updated signal, then another ranked display. Earlier recommendations influence which interactions can occur, and those interactions can reinforce the next recommendations.
This is one possible feedback pattern, not proof that all systems trap all users or that a mystery reader must dislike history. Clicking only among displayed items tells us little about unseen alternatives. Repeating the same inferred interest more confidently does not turn it into a verified statement about a person's identity or learning needs.
Add a visible diversity constraint
Now apply a rule that allows at most two items of the same genre in a three-item display, keeping the score order while scanning candidates. In every simulated round, choose B1 and B2, skip B3 as the third mystery item, and include B4 next. The display is B1, B2, B4, which exposes a history title. It does not prove the learner wants or will learn from that title.
A maximum of two per genre is a declared design choice, not a universal definition of diversity. It does not guarantee balanced viewpoints, reliable sources or accessible reading levels. If the catalogue contains only mystery items, the rule returns two rather than inventing a history book or silently breaking its own cap. Explain such boundaries to users instead of hiding them.
Execute or trace the complete ranking reference
The following original Python program was actually executed during course preparation. It reproduces the base ranking and three controlled rounds. It trains no model, calls no service and stores no real history. Learners can trace it on paper or use a teacher-approved environment.
def rank(items, profile=None, cap=None, limit=3):
profile = {} if profile is None else profile
if type(profile) is not dict or any(type(v) is not int or v < 0
for v in profile.values()):
raise ValueError("profile counts must be nonnegative integers")
if type(limit) is not int or limit < 0:
raise ValueError("limit must be a nonnegative integer")
if cap is not None and (type(cap) is not int or cap < 1):
raise ValueError("cap must be a positive integer")
if any(len(r) != 4 or type(r[3]) is not int or r[3] < 0 for r in items):
raise ValueError("items need ID title genre and nonnegative base count")
if len({r[0] for r in items}) != len(items):
raise ValueError("item IDs must be unique")
ordered = sorted(items, key=lambda r: (-(r[3] + 2 * profile.get(r[2], 0)), r[0]))
selected = []
per_genre = {}
for row in ordered:
if len(selected) == limit:
break
genre = row[2]
if cap is not None and per_genre.get(genre, 0) >= cap:
continue
selected.append(row)
per_genre[genre] = per_genre.get(genre, 0) + 1
return selected
catalog = [
("B1", "The Lantern Clue", "mystery", 8),
("B2", "The Clocktower Note", "mystery", 7),
("B3", "The Missing Map", "mystery", 6),
("B4", "Letters from an Old Harbour", "history", 3),
("B5", "Small Experiments Big Questions", "science", 2),
("B6", "Rain on the Window", "poetry", 1),
]
print("popularity_only", [r[0] for r in rank(catalog)])
profile = {"mystery": 2}
for round_number in range(1, 4):
plain = rank(catalog, profile)
diverse = rank(catalog, profile, cap=2)
print("round", round_number, "mystery_count", profile["mystery"],
"plain", [r[0] for r in plain], "diverse", [r[0] for r in diverse])
clicked_genre = plain[0][2]
profile = {**profile, clicked_genre: profile.get(clicked_genre, 0) + 1}Expected popularity-only output is B1 B2 B3. Rounds one to three show mystery counters two, three, four before display; plain lists B1 B2 B3 and diverse lists B1 B2 B4. The counter ends at five after the last prepared click. Empty catalogues or a display limit of zero produce no items. Invalid counts, caps and duplicate item IDs are rejected instead of silently inventing a ranking.
Change one signal with the rest fixed
Return to the original profile mystery_count two. In a separate comparison, change only B4's base count from three to thirteen. B4 now scores thirteen and ranks ahead of B1 twelve and B2 eleven. The top three becomes B4 B1 B2. The profile, genre labels and scoring formula stayed fixed; one signal changed the order.
Do not claim that the higher fictional click count makes B4 more historically accurate. Ranking and source evaluation ask different questions. Retain the original and changed catalogue versions and state which one supplied each result. Try the reranking before reading the worked checks; a partial step is B4's new score thirteen plus twice zero, which is thirteen.
Give people control and evaluate a different need
Compare three kinds of user control
An explanation can say which available signals influenced a suggestion. For B1 in the first round, the base is eight and the mystery contribution four, giving twelve. That is an explanation of this score, not proof the recommendation is suitable. A real product may need more information to explain an outcome responsibly.
A reset can clear the prepared genre counter. With an empty profile, this program returns the original base ranking B1 B2 B3. Reset does not guarantee varied genres because the frozen popularity list itself is dominated by mysteries. Combine appropriate controls rather than promising that a reset solves every problem. This worksheet reset is not a real account-data deletion request or a change to MOOCwiki settings.
An explicit goal or browse control can let a learner choose history directly. In our original catalogue B4 is the only history item, so a history-only browse yields B4, without pretending to fill three places. A diversity cap changes exposure but does not replace the learner's stated need. A teacher can help evaluate suitability, source reliability and reading accessibility separately from the score.
CodeAI: AI: Training Data & Bias. Ask which interests or materials may be absent and whether a system makes them harder to discover. The video does not establish that our cap ensures fairness or that a particular real recommendation service is safe. Keep claims tied to the actual signals, options and evidence inspected.
Test boundaries and keep uncertainty
Check the original score tie rule by giving two items the same score: item ID decides, not an undisclosed random choice. Reverse the input catalogue and verify the ranked IDs remain the same. Check a cap of one: the original personalised display becomes B1 B4 B5, representing three genres. With only the three mystery candidates and cap two, only B1 B2 are returned.
These finite tests check implementation behaviour under declared rules. They do not measure a real learner's satisfaction, truth preferences or long-term learning. A future trial in which someone explains why they clicked an unusual item is proposed but has not been performed here. Do not infer a permanent dislike from an unclicked item that was never displayed.
Transfer to a different information need
For a fictional science-information task, relevance to the question and trustworthy evidence may matter more than overall clicks. Redefine candidate sources, goal, criteria and controls. A popular video title or familiar logo alone is not evidence that its scientific claim is correct. Check the underlying source and date, and ask an appropriate adult when uncertain.
Use only this synthetic catalogue and profile. A classroom does not need to collect identifiable browsing histories to learn the principle. A separate live tool requires school approval, age-appropriate conditions, supervision and suitable privacy safeguards. Provide a non-digital equivalent and let learners state interests without making them public by default.
Keep the original catalogue, score rule, three round logs, one-signal change, cap/reset/browse results and limitations. State two limits: clicks are incomplete signals and the candidate pool constrains exposure. Propose source checking, explicit goals and meaningful user control as responses. A useful recommendation should be open to challenge; no quiz score proves that a system knows someone's true preferences. This is editorial AI Year 6 enrichment linked to effective information use.
Interactive Tasks
Quiz: Test Your Knowledge
What does a click establish in this lesson? (An observed or prepared interaction with several possible explanations) (!A permanent complete preference) (!A verified statement about every learner) (!The truth of a book content claim)
What is the initial personalised score for B1? (Twelve) (!Eight) (!Four) (!Two)
Which original items lead the popularity-only list? (B1 B2 B3) (!B4 B5 B6) (!B1 B4 B5) (!B6 B5 B4)
What happens to the mystery counter after three prepared clicks? (It rises from two to five) (!It stays at two) (!It becomes a real pupil record) (!It proves history is disliked)
Which display results from a maximum of two per genre? (B1 B2 B4) (!B1 B2 B3) (!B4 B5 B6) (!B1 B3 B2)
What is the third-round mystery counter before display? (Four) (!Five) (!Two) (!Three)
What changes when only B4 base count becomes thirteen? (The top three becomes B4 B1 B2) (!B4 automatically becomes more trustworthy) (!The learner profile is erased) (!The scoring formula must secretly change)
Why may a reset still show three mysteries? (The frozen base catalogue ranking also favours mysteries) (!The reset always guarantees equal genres) (!The program knows a permanent preference) (!History books are forbidden)
What does an empty catalogue produce? (No recommendations) (!Three invented books) (!A real account upload) (!A guaranteed history title)
Which check is separate from ranking popularity? (Evaluate source reliability and suitability for the current need) (!Assume the top item is always accurate) (!Treat an unshown item as disliked) (!Infer identity from repeated clicks)
Memory Game
| Content feature | A description of an item such as genre |
| User signal | An interaction or explicitly supplied interest |
| Ranking score | A value used to order candidates under a chosen rule |
| Feedback loop | Earlier outputs shape signals that influence later outputs |
| Diversity cap | A declared limit on repeated categories in the display |
| Reset | Clear the simulated profile contribution without changing the base catalogue |
Drag and Drop
| An unexplained repeated mystery suggestion | Inspect the signals and allow a stated new goal |
| A click on the only visible topic | Avoid treating exposure-limited behaviour as a complete preference |
| An item with the highest click score | Check suitability and reliability separately |
| A cap with too few eligible genres | Return fewer items instead of fabricating candidates |
| A history-information need | Use explicit relevance and appropriate source checking |
Crossword Puzzle
| GENRE | What content category labels mystery history or poetry? |
| SIGNAL | What observed or supplied clue influences ranking? |
| SCORE | What value orders candidates under the chosen rule? |
| FEEDBACK | What returns from earlier interaction to affect later results? |
| RESET | What clears the simulated profile counter? |
| CHOICE | What should remain meaningful for the learner? |
LearningApps
Optional extra practice. Choose an appropriate activity with your teacher. If the external collection is unavailable, use the native activities and independent tasks in this course.
Cloze Text
Open Tasks
Easy
- Explain one signal: Give two possible reasons for a click without inferring permanent identity.
- Calculate one score: Combine B1 base eight and mystery counter two using the stated rule.
- Trace a round: Record the display, prepared click and before/after counter.
- Read a control result: Explain why cap two includes B4 without proving a history preference.
Standard
- Simulate all three rounds: Keep base counts fixed and record scores, displays and counters.
- Compare ranking modes: Contrast popularity-only, personalised and capped exposure.
- Change one signal: Hold the original profile fixed and rerank after changing only B4 base.
- Test reset and browse: Compare an empty profile, cap one and an explicit history goal.
Advanced
- Challenge an inferred preference: Explain how exposure constrains evidence and why unshown items supply little information.
- Test scarce candidates: Use a mystery-only catalogue and keep the cap without inventing extra items.
- Redesign for science information: Specify new relevance and source-checking criteria for a different need.
- Keep an evidence and control log: Record the finite tests, two limitations, a proposed unperformed trial and meaningful learner/teacher options.
Learning Check
- Initial scores and signals: Calculate and interpret the first ranking.
- Three-round feedback loop: Trace the controlled exposure and updates.
- Diversity and candidate limits: Explain cap results and their limits.
- Controlled signal change: Rerank with only one base signal changed.
- Explanation reset and goal: Compare user control options.
- Transfer and evidence limits: Design a different information-use task.
Worked checks and assessment guidance
- Original base counts are eight, seven, six, three, two, one. With mystery counter two, add four to mystery items: B1/B2/B3 score twelve/eleven/ten. These are arbitrary ranking scores, not probabilities, truth ratings or verified permanent preferences.
- Before displays the mystery counter is two, three, four. The plain list remains B1 B2 B3, with scores12/11/10 then14/13/12 then16/15/14. A supplied click on first item B1 increments after each round, ending at five. No real user behaviour was observed.
- Cap two yields B1 B2 B4; cap one yields B1 B4 B5 with the original profile. Only three mystery candidates plus cap two yields two items B1 B2. The rule changes exposure, not proof of interest, balanced viewpoints or reliable content.
- Restore mystery counter two. B4 base thirteen gives score thirteen, followed by B1 twelve and B2 eleven, so B4 B1 B2. Other fields and formula are unchanged. Preserve versions and do not infer historical accuracy from more clicks.
- Explain score components. Resetting the profile removes its contribution but leaves the base list B1 B2 B3, so it need not restore genre variety. A history-only browse yields the one history item B4. None of these worksheet controls changes real account data or replaces suitability/source checking.
- For science information, redefine relevance, source reliability, candidate pool and user control. Clicks are incomplete and exposure-limited signals. Keep the finite log, allow explicit goals and teacher review, avoid identifiable browsing data and label the proposed explanation-of-click trial unperformed. A top rank is not an accuracy guarantee.
Learning evidence
Retain the synthetic catalogue and original profile, three rounds with before/after counters, exact score rule, capped displays, one-signal change and reset/browse boundary checks. Explain incomplete signals and exposure limits, meaningful controls, a changed information need and a genuinely proposed unperformed trial. No real preferences, user tracking, trained recommender or account-setting changes are established.
OER and Wikipedia
DfE Computing supplies the KS2 search/information-use anchor; this annual AI sequence is editorial. Google: recommendations overview and Google: content-based filtering inform the conceptual distinction between item features and interaction signals. The fictional books, scoring rule, controlled clicks, constraints and program are original, not publisher examples, a trained service or a real person's history. Reviewed 15 November 2026.
International books at Kent Library.jpg, Roc0ast3r, CC0. Unchanged library photograph; contextual illustration, not a source for invented titles or click counts.
https://www.youtube.com/watch?v=OeU5m6vRyCk — CodeAI: AI: What is Machine Learning? Original recording embedded unchanged.
https://www.youtube.com/watch?v=x2mRoFNm22g — CodeAI: AI: Training Data & Bias. Original recording embedded unchanged. Teacher-supported context; no live pupil account or tracking required.
Linked learning areas
aiMOOC Projects
Schulfach+


aiMOOCs


aiMOOC Projekte


NEWSLernweltNOAH fragen