searchsim/cognitive-traces-movielens

---
license: cc-by-4.0
task_categories:
- text-classification
- token-classification
language:
- en
tags:
- information-retrieval
- user-simulation
- cognitive-modeling
- information-foraging-theory
- search-logs
pretty_name: "Cognitive Traces — MovieLens"
size_categories:
- 100K<n<1M
---

# 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

```python
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

```bibtex
@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](https://traces.searchsim.org/)
- [GitHub Repository](https://github.com/searchsim-org/cognitive-traces)
- [Annotation Tool](https://github.com/searchsim-org/cognitive-traces)

license: CC BY 4.0
task_categories:
- 文本分类（text-classification）
- 令牌分类（token-classification）
language:
- 英语（en）
tags:
- 信息检索（information-retrieval）
- 用户模拟（user-simulation）
- 认知建模（cognitive-modeling）
- 信息觅食理论（information-foraging-theory）
- 搜索日志（search-logs）
pretty_name: "认知轨迹 — MovieLens"
size_categories:
- 100K<n<1M

# 认知轨迹 — MovieLens

## 数据集描述

本数据集为MovieLens数据集配套提供**认知轨迹标注（cognitive trace annotations）**，由下述文献中提及的多智能体标注框架生成：

> **《超越点击：搜索场景下认知轨迹的推断框架》** Saber Zerhoudi、Michael Granitzer. ECIR 2026.

每个用户事件（评分、信念征询、信念预测）均标注有来自**信息觅食理论（Information Foraging Theory, IFT）**的认知标签，同时附带完整的标注链路（分析师、评论者、审定者）与置信度分数。

## 数据集统计

| 指标 | 数值 |
|--------|-------|
| 会话数 | 10,274 |
| 事件数 | 111,561 |
| 动作类型 | RATE、BELIEF_ELICIT、BELIEF_PREDICT（共3类） |
| 认知标签 | 6类，分别为FollowingScent、ApproachingSource、ForagingSuccess、DietEnrichment、PoorScent、LeavingPatch |

## 快速上手

python
from datasets import load_dataset

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

# 访问数据集
print(ds["train"][0])

# 按认知标签筛选
struggling = ds["train"].filter(lambda x: x["cognitive_label"] == "PoorScent")
print(f"PoorScent标签事件数：{len(struggling)}")

# 获取指定会话的所有事件
session = ds["train"].filter(lambda x: x["session_id"] == "ml_42170_0")
for event in session:
print(f" {event["action_type"]}: {event["cognitive_label"]}")

## 列结构

| 列名 | 数据类型 | 描述 |
|--------|------|-------------|
| `session_id` | 字符串 | 唯一会话标识符 |
| `event_id` | 字符串 | 唯一事件标识符 |
| `event_timestamp` | 字符串 | ISO格式时间戳 |
| `action_type` | 字符串 | 用户动作类型（RATE、BELIEF_ELICIT、BELIEF_PREDICT） |
| `content` | 字符串 | 包含movie_id、movie_title、genres与rating的JSON数据 |
| `cognitive_label` | 字符串 | 最终的信息觅食理论认知标签 |
| `analyst_label` | 字符串 | 分析师代理提出的标签 |
| `analyst_justification` | 字符串 | 分析师的标注理由 |
| `critic_label` | 字符串 | 评论者代理提出的标签 |
| `critic_agreement` | 字符串 | 评论者是否同意分析师的标注 |
| `critic_justification` | 字符串 | 评论者的标注理由 |
| `judge_justification` | 字符串 | 审定者的最终决策理由 |
| `confidence_score` | 浮点型 | 框架置信度（取值范围0–1） |
| `disagreement_score` | 浮点型 | 分析师与评论者的分歧度（取值范围0–1） |
| `flagged_for_review` | 布尔型 | 是否标记为需人工复核 |
| `pipeline_mode` | 字符串 | 标注流水线模式 |

## 信息觅食理论认知标签

| 标签 | 信息觅食理论概念 | 含义解释 |
|-------|-------------|----------------|
| FollowingScent | 信息追踪（Information scent following） | 用户沿着优质线索开展探索 |
| ApproachingSource | 源接近（Source approaching） | 用户逐步靠近目标信息 |
| ForagingSuccess | 成功觅食（Successful foraging） | 用户找到所需信息 |
| DietEnrichment | 信息拓展（Diet enrichment） | 用户拓宽信息摄入范围 |
| PoorScent | 劣质线索（Poor information scent） | 线索质量下降 |
| LeavingPatch | 放弃搜索域（Patch leaving） | 用户放弃当前搜索方向 |

## 源数据集

本数据集基于MovieLens-25M（Harper与Konstan, TiiS 2016）构建，包含电影评分与信念征询、预测事件。`content`字段包含带有movie_id、movie_title、genres与rating的JSON数据。

## 引用格式

bibtex
@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}
}

## 许可证

CC BY 4.0。本数据集的认知标注基于知识共享署名4.0（Creative Commons Attribution 4.0）协议发布。底层源数据集拥有各自的许可证，请参阅原始数据集提供商的相关说明。

## 相关链接
- [论文](https://traces.searchsim.org/)
- [GitHub仓库](https://github.com/searchsim-org/cognitive-traces)
- [标注工具](https://github.com/searchsim-org/cognitive-traces)