DecisionTrees-AICFT-Frames: CODAP Arbor Screen-Recording Dataset

This dataset contains 5,825 behaviour-coded screen-recording frames from CODAP Arbor sessions conducted as part of a decision tree learning activity. CODAP Arbor is an interactive, browser-based machine learning environment in which learners construct and evaluate decision tree classifiers on tabular datasets. Fifteen pre-service teachers were recorded while working in CODAP Arbor; frames were selected automatically from their screen recordings using motion and speech triggers, then annotated by a single researcher with behavioural codes, screen-context labels, and confidence ratings.

The dataset supports research on observable learning behaviours during data-driven modelling tasks. It is distinct from the companion research archive, which contains worksheets, assessment rubrics, Python notebooks, and additional project materials.

---

Dataset at a glance

Property	Value
Dataset type	Annotated screen-recording frames
Research domain	Multimodal learning analytics (MMLA), pre-service teacher education
Participants	15 (pseudonymised pre-service teachers)
Sessions	2 CODAP Arbor screen-recording sessions
Extracted frames	5,825
Primary behaviour categories	9
Data modalities	Screen images (JPG), per-frame annotation metadata, platform event log
Annotation approach	Single-researcher behavioural coding; confidence recorded per frame
Instruction language	Turkish
Annotation language	English
Image format	JPEG, max width 1920 px
Repository size	~1.8 GB
License	CC BY 4.0

---

Research context and intended use

Participants engaged in a supervised classroom activity in which they used CODAP Arbor to construct decision tree classifiers on provided datasets. The activity was designed to develop competencies aligned with the UNESCO AI Competency Framework for Teachers (AI-CFT). Screen recordings captured each participant's on-screen activity throughout the session.

This dataset may support research questions such as:

• How do pre-service teachers distribute their time across observable modelling sub-tasks?
• What screen contexts co-occur with specific behavioural codes?
• How does on-screen activity evolve across task phases?
• How can automated frame selection be combined with behavioural annotation in classroom research settings?

Interpretive caution. A coded frame documents observable screen activity at a sampled moment. It is not a direct measure of a participant's cognitive state, understanding, or competence. Behavioural codes reflect the researcher's interpretation of visible activity and should not be treated as ground-truth evidence of learning.

---

Data collection and frame-selection procedure

Screen recordings were made for each participant across two sessions. Frames were selected from continuous video using an automated multi-trigger extraction pipeline with the following selection criteria:

Trigger code	Description
motion_threshold_exceeded	Screen changed by more than 1.2% of pixels between consecutive frames, with usable audio available
motion_only_fallback	Screen changed by more than 0.8% of pixels in recordings without usable audio
speech_anchor_midpoint	A speaker was detected at the midpoint of a speech segment
codap_static_gap_fill	No frame had been captured for 90 seconds on the CODAP task screen
silent_screen_gap_fill	No frame had been captured for 90 seconds in a silent recording

Additional processing steps were applied before a frame was retained: visual duplicates were removed using perceptual hashing (pHash threshold: 4); frames that matched non-task or calibration screen templates were discarded; a 1-second cooldown prevented burst extraction during rapid transitions.

The screen_change_pct column records the pixel-change percentage at the moment of extraction. It is populated for motion-triggered frames and empty for speech and gap-fill frames.

Timestamps (timestamp_s) refer to elapsed seconds within the participant's own screen recording. They are not wall-clock times and cannot be used to determine the absolute time of day.

Published images are full extracted frames (not cropped thumbnails) at the original capture resolution, up to a maximum width of 1920 pixels.

Sampling note. Selected frames represent moments judged informative by the extraction pipeline, not a continuous or uniform record. Many on-screen actions occur between extracted frames and are not represented in this dataset.

---

Behavioural annotation

Each frame was annotated by a single researcher. No inter-rater reliability procedure is documented for this dataset.

Primary behaviour categories

Code	Description
EXPLORE_DATA	Participant is examining attributes, distributions, or sample rows in the data table
SELECT_TARGET	Participant is choosing or changing the target (dependent) variable
BUILD_TREE	Participant is actively constructing the decision tree structure
TUNE_THRESHOLD	Participant is adjusting a split threshold or node parameter
EVALUATE_MODEL	Participant is inspecting model performance metrics for a single model
COMPARE_MODELS	Participant is comparing performance across two or more models
INTERPRET_RESULTS	Participant is reading or discussing the meaning of a model result
IDLE_THINKING	No visible on-screen action; participant appears to be paused or reading
OFF_TASK	On-screen activity is unrelated to the decision tree task

Secondary behaviour

The behavior_secondary column records a second concurrent behaviour when two were clearly observable. It uses the same code vocabulary as the primary field. It is empty for the majority of frames.

Confidence and evidence

coder_confidence records the researcher's self-assessed confidence in the assigned primary behaviour label (LOW, MEDIUM, or HIGH). It reflects annotation certainty, not annotation correctness.

evidence_flags is a comma-separated list of specific on-screen observations that informed the annotation (for example, tree_has_nodes, accuracy_visible, overfitting_awareness). The field is empty when no specific evidence was recorded.

---

File structure and data dictionary

DecisionTrees-AICFT-Frames/
├── images/                        # 5,825 JPEG frames
├── metadata.csv                   # One row per frame; links each image to its annotation
├── codap_log_anonymized.csv       # CODAP Arbor platform event log (pseudonymised)
└── README.md                      # This file

Linking images to metadata

Each image is referenced by its file_name field in metadata.csv (e.g., images/Amy_codap_21apr_frame_0001.jpg). Image filenames use the original recording session codes (codap_21apr, codap_28apr); the session column in metadata uses the mapped identifiers session_1 and session_2.

metadata.csv columns

Record identity and provenance

Column	Description
file_name	Relative path to the image file (used by the dataset viewer to link images to rows)
student	Participant pseudonym
session	Session identifier: session_1 (21 April) or session_2 (28 April)

Temporal information and frame selection

Column	Description
timestamp_s	Elapsed seconds in the recording at the annotated frame
selection_trigger	Trigger code that caused the frame to be selected (see trigger table above)
screen_change_pct	Proportion of pixels that changed at extraction, as a percentage; empty for non-motion triggers
speaker	Speaker role detected at this moment: teacher, student, or empty

Behavioural annotations

Column	Description
behavior	Primary coded behaviour (see behaviour category table above)
behavior_secondary	Secondary behaviour, if a second was clearly observable; otherwise empty

Annotation confidence and evidence

Column	Description
coder_confidence	Annotation confidence: LOW, MEDIUM, or HIGH
evidence_flags	Comma-separated on-screen observation flags that informed the annotation

codap_log_anonymized.csv

The CODAP log records platform events (e.g., attribute selection, model runs, session events) generated by the CODAP Arbor application. Columns include an event identifier, creation timestamp, pseudonymised student identifier, action type, event timestamp in milliseconds, parameters, and raw data fields. The log covers events from both sessions.

The log is an independent event stream. It is not aligned frame-by-frame with the screen recordings. Platform events and extracted frames share only the participant pseudonym and session as linking keys; temporal alignment requires matching timestamp_s from the frame metadata against timestamp_ms from the log.

---

Data quality and limitations

• Temporal sampling. Extracted frames represent informative moments as judged by the pipeline, not every second of activity. On-screen events that do not trigger any extraction criterion are absent from the dataset.
• Single-researcher annotation. All behavioural codes were assigned by one researcher. No inter-rater reliability statistic is available. Labels reflect interpretive judgement and should be treated accordingly.
• Inference limits. Coded behaviours describe observable screen activity. They do not provide direct access to participants' cognitive processes, intentions, or learning outcomes.
• Frame-selection sensitivity. The motion threshold (1.2% pixel change) and deduplication settings mean that slow or visually subtle interactions may be underrepresented. Rapid screen transitions may produce multiple closely spaced frames.
• Missing annotation fields. A small number of frames have empty behavior, screen_context, or phase fields. These are included in the dataset but should be excluded from analyses that require complete annotation.
• CODAP log scope. The platform log covers both recorded sessions but does not have a verified one-to-one correspondence with the extracted frames.

---

Privacy, ethics, and responsible reuse

Participants are identified in this dataset by pseudonyms (Amy, Bruno, Helena, Iris, Irma, Isabel, Marco, Marcus, Melinda, Nadia, Serena, Shana, Sheila, Ulysses, Zara). Real names, institutional identifiers, and student identifiers have been removed from metadata fields.

Important limitations of the pseudonymisation approach. Pseudonymisation is not equivalent to full anonymisation. Screen images may contain contextual visual information — visible application content, on-screen text, or interaction patterns — that could, in combination with other information, reduce practical privacy protections. Users of this dataset should not attempt to identify participants or combine the dataset with external sources for that purpose.

Reuse guidance. This dataset is released for research purposes. Users are responsible for complying with applicable privacy and data protection regulations in their jurisdiction. The open license does not override any consent, ethics, or data-sharing obligations that may apply.

Note to users. Formal documentation of ethics approval, participant consent for open data sharing, and the scope of image de-identification has not been published alongside this dataset. Researchers who require such documentation before use should contact the dataset author before incorporating this data into their work.

---

License and reuse

This dataset is released under Creative Commons Attribution 4.0 International (CC BY 4.0). You are free to share and adapt the material for any purpose, provided you give appropriate credit.

Attribution requirement: cite the dataset as below and include the repository URL. An open license does not override applicable privacy or consent obligations.

---

Citation

Dede, M. (2026). DecisionTrees-AICFT-Frames. Hugging Face. https://huggingface.co/datasets/dedemerve/DecisionTrees-AICFT-Frames

@dataset{dede2026decisiontrees,
  author    = {Dede, Merve},
  title     = {DecisionTrees-AICFT-Frames},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/dedemerve/DecisionTrees-AICFT-Frames}
}

---

Related resources

Resource	URL
Companion research archive (worksheets, rubrics, notebooks, platform logs)	https://dedemerve.github.io/DecisionTrees-AICFT-Assessor/
Source code repository	https://github.com/dedemerve/DecisionTrees-AICFT-Assessor