Explicit and Implicit Feedback
Explicit feedback asks for a judgment, such as a rating or dislike. Implicit feedback observes behavior, such as an impression, click, save, completion, repeat play, or dwell time. Both are measurements produced by a particular interface and population.
Prerequisites
You should understand the event grain in users, items, context, and feedback.
Learning goals
- Distinguish a preference label from evidence about preference.
- State why non-interaction is not negative feedback.
- Preserve exposure when interpreting implicit behavior.
- Choose an aggregation and confidence rule that matches the objective.
Explicit is sparse; implicit is ambiguous
A five-star rating is a direct response to a prompt, but it may be sparse and selected: people choose which items to rate. A click is plentiful, but it can reflect curiosity, position, a misleading title, or accidental input. A completion may better match learning progress, yet still depends on item length and opportunity.
For implicit data, distinguish a binary preference proxy from confidence. Hu, Koren, and Volinsky define a preference indicator from observed consumption and a confidence that grows with evidence. This does not reveal a true hidden preference; it states how the training objective weights observed and unobserved pairs.
Worked example: preserve the event sequence
import csv
from pathlib import Path
path = Path("tests/fixtures/recommender-systems/interactions.csv")
with path.open(encoding="utf-8", newline="") as handle:
rows = list(csv.DictReader(handle))
u1_i1 = [
row["event_type"] for row in rows
if row["user_id"] == "u1" and row["item_id"] == "i1"
]
u1_i2 = [
row for row in rows
if row["user_id"] == "u1" and row["item_id"] == "i2"
]
action_types = {"click", "save", "complete", "rating"}
exposed_without_action = (
any(row["event_type"] == "impression" for row in u1_i2)
and not any(row["event_type"] in action_types for row in u1_i2)
)
print(f"u1-i1 events={u1_i1}")
print(f"u1-i2 exposed_without_action={exposed_without_action}")
Output:
u1-i1 events=['impression', 'click', 'complete', 'rating']
u1-i2 exposed_without_action=True
Even u1's exposed-but-unclicked i2 is not a clean dislike. The learner may not have examined rank 2, may have run out of time, or may already know the topic. It is evidence under a logged exposure, not a universal negative label.
Define relevance before aggregation
For a completion objective, a practical binary relevance definition might be “a completion in the seven-day test horizon.” For an immediate-start objective, it might be “a click during the visit.” Ratings could define graded relevance, for example 1–5, but only among users and items that receive ratings. Record the definition and horizon beside every result.
When repeated actions exist, choose deliberately: latest rating, maximum rating, any completion, count capped at one per request, or a decayed count. Never let an accidental duplicate event silently become extra preference.
Limitations and common mistakes
- Missing feedback mixes no exposure, no attention, no opportunity, and possible dislike.
- A stronger action is not always more preference; repeated plays can indicate difficulty.
- Explicit ratings are affected by who chooses to rate and by the scale's interpretation.
- Combining actions into weights creates a product-specific target, not a universal preference score.
- Negative sampling changes the learning and evaluation task; sampled items must be documented.
Practice
For clicks, saves, completions, and ratings, write one plausible preference interpretation and one competing explanation. Then define relevance for a top-3 evaluation and state which event duplicates count once.
Next steps
Next, separate exposure, position, and selection bias from preference.
Primary sources
Checked 2026-10-17: Hu, Koren, and Volinsky, “Collaborative Filtering for Implicit Feedback Datasets”, and Joachims, Swaminathan, and Schnabel, “Unbiased Learning-to-Rank with Biased Feedback”.