searchsim/cognitive-traces-movielens

Cognitive Traces — MovieLens

Dataset Description

This dataset contains cognitive trace annotations for the MovieLens dataset, produced by the multi-agent annotation framework described in:

Beyond the Click: A Framework for Inferring Cognitive Traces in Search Saber Zerhoudi, Michael Granitzer. ECIR 2026.

Each user event (rating, belief elicitation, belief prediction) is annotated with a cognitive label from Information Foraging Theory (IFT), along with the full annotation chain (analyst, critic, judge) and confidence scores.

Dataset Statistics

Metric	Value
Sessions	10,274
Events	111,561
Action Types	RATE, BELIEF_ELICIT, BELIEF_PREDICT (3)
Cognitive Labels	6 (FollowingScent, ApproachingSource, ForagingSuccess, DietEnrichment, PoorScent, LeavingPatch)

Quick Start

from datasets import load_dataset

ds = load_dataset("searchsim/cognitive-traces-movielens")

# Access the data
print(ds["train"][0])

# Filter by cognitive label
struggling = ds["train"].filter(lambda x: x["cognitive_label"] == "PoorScent")
print(f"Events with PoorScent: {len(struggling)}")

# Get all events for a session
session = ds["train"].filter(lambda x: x["session_id"] == "ml_42170_0")
for event in session:
    print(f"  {event["action_type"]}: {event["cognitive_label"]}")

Column Schema

Column	Type	Description
session_id	string	Unique session identifier
event_id	string	Unique event identifier
event_timestamp	string	ISO timestamp
action_type	string	User action type (RATE, BELIEF_ELICIT, BELIEF_PREDICT)
content	string	JSON with movie_id, movie_title, genres, and rating
cognitive_label	string	Final IFT cognitive label
analyst_label	string	Analyst agent's proposed label
analyst_justification	string	Analyst's reasoning
critic_label	string	Critic agent's proposed label
critic_agreement	string	Whether Critic agreed with Analyst
critic_justification	string	Critic's reasoning
judge_justification	string	Judge's final decision reasoning
confidence_score	float	Framework confidence (0–1)
disagreement_score	float	Analyst–Critic disagreement (0–1)
flagged_for_review	bool	Whether flagged for human review
pipeline_mode	string	Annotation pipeline mode

IFT Cognitive Labels

Label	IFT Concept	Interpretation
FollowingScent	Information scent following	User pursuing a promising trail
ApproachingSource	Source approaching	User converging on target information
ForagingSuccess	Successful foraging	User found desired information
DietEnrichment	Diet enrichment	User broadening information intake
PoorScent	Poor information scent	Trail quality deteriorating
LeavingPatch	Patch leaving	User abandoning current direction

Source Dataset

Based on MovieLens-25M (Harper & Konstan, TiiS 2016). Contains movie ratings and belief elicitation/prediction events. Content field includes JSON with movie_id, movie_title, genres, and rating.

Citation

@inproceedings{zerhoudi2026beyond,
  title={Beyond the Click: A Framework for Inferring Cognitive Traces in Search},
  author={Zerhoudi, Saber and Granitzer, Michael},
  booktitle={Proceedings of the 48th European Conference on Information Retrieval (ECIR)},
  year={2026}
}

License

CC-BY-4.0. The cognitive annotations are released under Creative Commons Attribution 4.0. The underlying source datasets have their own licenses — please refer to the original dataset providers.

Links

• Paper
• GitHub Repository
• Annotation Tool