This notebook is a work in progress. I'm currently focusing on exploratory data analysis and will be extending it with preprocessing steps and model development soon.
\nOkay, let me walk through what I’m seeing…
\n\n⏩ Core Tables & Their Scales:
\ntrain.csv (~575K rows × 341 cols): \n A massive stream of per-timestamp readings—implies a lot of granular motion data but only about ~5K unique sequence_id groups.train_demographics.csv (81 rows × 8 cols): \n Profiles for each participant—only 81 unique subjects despite hundreds of thousands of rows in train.csv.test.csv (107 rows × 336 cols): \n A tiny snippet of sensor streams—means the test set only exposes a few full-sequence dumps, so look out for distributional shifts.test_demographics.csv (2 rows × 8 cols): \n Just two participants—extremely limited demographic coverage in the hold-out.⏩ Data Cleanliness & Duplicates:
\n-1 or missing values for Time-of-Flight (ToF) gracefully—probably impute or mask.⏩ Participant & Contextual Insights:
\nsequence_counter in train.csv spans from 0 up to ~700, with a median around 35—some gestures are very long, others are brief. \n - In test.csv, the max sequence_counter is ~55, so test gestures tend to be shorter.⏩ Sensor-Based Observations:
\n🌟 Key Practical Takeaways:
\n🗣 Additional Strategic Reflections:
\nFeature Engineering (FE) and avoid overly complex ensembles—simple models with good FE often outperform large stacks here.sequence_id × gesture) carefully.1️⃣ Sequence Length (sequence_counter) ⏱️
✅ Next Steps: Create sequence‐agnostic summaries (e.g., means, extremes) or apply dynamic models (RNN/LSTM with masking).
\n2️⃣ Age (age) 👶👴
✅ Next Steps: Encode age as continuous or bin into child vs. adult flags. Use caution when adding age interactions, since only two demo points appear in test.
\n3️⃣ Height (height_cm) 📏
135 cm) and higher extremes (~190 cm).168 and 174 cm) are squarely within the train range, so no major “height gap” is present for test runs.✅ Next Steps: Normalize accelerometer features by limb length or include height as an auxiliary covariate—but avoid heavy transformation since distribution is fairly symmetric.
\n4️⃣ Shoulder‐to‐Wrist Length (shoulder_to_wrist_cm) 🤸
41–71 cm with most participants in the 49–55 cm range. This roughly correlates with height, but captures the specific leverage of motion. 52 cm, which is near the median. No extreme arm length mismatch between train and test.✅ Next Steps: Consider scaling positional features by shoulder_to_wrist_cm to normalize gesture amplitude.
If ToF grids are sensitive to distance, arm length scaling may stabilize proximity features.
5️⃣ Elbow‐to‐Wrist Length (elbow_to_wrist_cm) 💪
18–44 cm) but still variable. This measurement captures forearm length—directly linked to how far the wrist can travel when scratching or pulling hair near the face.25 cm) are near the train median. Again, no extreme train–test gap.✅ Next Steps: Similar to shoulder length, normalize accelerometer/rotation magnitudes by elbow_to_wrist_cm to account for per‐person leverage.
6️⃣ Accelerometer Magnitude (Aggregated – acc_mag_mean, acc_mag_std, acc_mag_max) 📈
acc_mag_mean is quite stable around ~10 m/s² for both train and test, indicating a baseline motion level when gestures occur. acc_mag_std shows that train sequences can vary widely—some are nearly static, others very jerky. Test sequences cluster in a narrower band of variability. acc_mag_max highlights the peak intensity of motion. Train has some extreme peaks (forceful tugs), whereas test peaks are more moderate—again a slight distributional shift. ✅ Next Steps: Derive features such as peak-to-mean ratios or threshold‐based flags—gestures above a certain acceleration threshold might be high‐confidence BFRBs. Account for the narrower test variance so the model doesn’t over‐penalize “normal” extremes.
7️⃣ Rotation Angle (Aggregated – rot_angle_mean, rot_angle_std, rot_angle_max) 🔄
rot_angle_mean centers around ~2.4 rad for both sets—indicative of wrist tilt involved in most gestures. Train vs. Test alignment is strong here, which is reassuring for model generalization.rot_angle_std is lower in test, meaning test participants move their wrist more uniformly (less jitter). rot_angle_max shows how far a wrist can twist at its peak. Train reaches near π (180°), test peaks are slightly lower (~2.7 rad). ✅ Next Steps: Create angular velocity features—differences in consecutive rot_angle could identify sharp twists vs. gentle movements. Use caution since test has narrower rotational extremes.
8️⃣ Thermopile Averages (Aggregated – thm_1_mean … thm_5_mean) 🌡️
thm_1_mean values (even near zero for a few train sequences) suggest times when the wrist was far from any warm surface—probably during “Transition” and “Pause” phases. thm_i_mean rising could pinpoint the exact wrist‐to‐face contact moment.thm_2_mean sometimes spikes in test (up to ~31–32 °C), likely due to closer proximity or warmer body parts (e.g., forehead). ✅ Next Steps: Compute delta_temperatures (e.g., current minus initial temp) to highlight warm‐contact events, ignoring static baseline differences between sequences.
9️⃣ Thermopile Variability (Aggregated – thm_1_std … thm_5_std) 📉
thm_*_std in train is noticeably higher (some sequences show up to ~5 °C fluctuation), reflecting dramatic flows between “far from skin” and “touching skin.” ✅ Next Steps: Use thm_i_std as a feature to detect “novelty” in thermal patterns. High std during the “Gesture” phase flags potential BFRBs.
🔟 ToF Averages (Aggregated – tof_1_mean … tof_5_mean) 📏
tof_i_mean spans the full 0–250 range, showing that some gestures move the wrist very close (distance near 0) while others stay farther away. tof_mean near zero indicates direct skin contact; higher means “air” gestures like waving or writing in the air.tof_mean values seem clustered mid‐range for some sensors (e.g., ~130–180), suggesting test gestures are often intermediate distances—maybe a lot of transitions or ambiguous proximity. ✅ Next Steps: Create proximity flags: e.g., is_close = (tof_i_mean < threshold). Use a lower threshold for face contact (e.g., < 20) and higher for “in‐air.”
1️⃣1️⃣ ToF Variability (Aggregated – tof_1_std … tof_5_std) 📈
tof_std shows large variation (some sequences flirting with ~140), indicating rapid in‐and‐out motions—typical of quick BFRB tugs. tof_std is generally lower (~20–40), so test gestures exhibit smoother proximity changes or more stable distances. tof_std heavier for train (more jittery patterns) but temper its impact when test std is uniformly low.✅ Next Steps: Engineer a feature like proximity_jitter = tof_std / tof_mean to capture “relative instability” in distance, making it robust to absolute sensor offsets.
1️⃣2️⃣ Phase (phase) 🚦
NaN), so we must derive our own phase segmentation or ignore it at inference. ✅ Next Steps: If using train phases, build separate pipelines for Transition vs. Gesture intervals. At test time, use heuristic (e.g., first few frames → transition).
1️⃣3️⃣ Behavior (behavior) 🤹
✅ Next Steps: Create a rule to tag “gesture”‐like rows in test: e.g., when acc_mag & thm_mean both spike, label as performing behavior.
1️⃣4️⃣ Orientation (orientation) 🪑🤸
is_lying) and encourage model invariance to it.✅ Next Steps: One-hot encode orientation in train, then train a small classifier (or clustering) to predict orientation in test—use predicted posture as feature.
\n1️⃣5️⃣ Sequence Type (sequence_type) 🎯❌
Target (BFRB-like) vs. 40% Non-Target. This class imbalance tells us our model must be robust to a slight majority of BFRBs but not over‐predict them. ✅ Next Steps: Use stratified cross‐validation to preserve true ratio. Monitor both binary F1 and gesture‐level F1 (macro) during tuning.
\n1️⃣6️⃣ Adult vs. Child (adult_child) 👶👨🦳
adult_child and sensor aggregates.✅ Next Steps: Create interaction terms (e.g., acc_mag_mean × adult_child) or build a small model to detect child vs. adult in test and adjust thresholds accordingly.
1️⃣7️⃣ Sex (sex) ♂️♀️
sex as a feature, but watch out for overfitting since test is only two participants (one male, one female).✅ Next Steps: Incorporate sex sparingly—perhaps as a one‐hot input to neural net so that model can learn subtle gender‐specific motion differences.
1️⃣8️⃣ Handedness (handedness) 🤚
✅ Next Steps: No special handling needed unless left-handers appear in new data; focus on edge cases like ambidextrous patterns.
\n1️⃣9️⃣ Gesture Labels (gesture) 🎭
Text on phone) and non-BFRBs (e.g., Drink from bottle) each comprising ~5–10% of data. Some low-frequency gestures (~Pinch knee/leg skin) appear in just ~1–3% of samples. NaN), so we rely entirely on sensor features. The model must generalize across these 18 classes, paying special attention to similar movements (e.g., Eyebrow – pull hair vs. Forehead – pull hairline). gesture1️⃣ Multimodal, Nearly Uniform Spread Across Many Classes 🧐
\nText on phone and Neck – scratch each occupy roughly 10% of the data, but lots of mid‐frequency gestures hover around 7–8% (e.g., Eyebrow – pull hair, Forehead – scratch).Write name on leg and Pinch knee/leg skin sit near the 1.5–2% range, suggesting a long tail of rarer events.✅ Practical Implications: \n - A model must handle a fine‐grained, multi‐class ratio without overfitting to the most common gestures. \n - The moderate class imbalance (10% vs. ~2%) signals a need for class‐balanced sampling or weighted loss so rare gestures still get learned. \n
\n2️⃣ All Gestures Are Potentially BFRB or Non‐BFRB—Labeling Matters 📑
\nForehead – pull hairline, Cheek – pinch skin), and half are non‐BFRB everyday actions (e.g., Drink from bottle/cup, Wave hello).✅ Practical Implications: \n - Start with a binary‐classification pipeline (BFRB vs. non‐BFRB). \n - Then refine to the multi‐class problem, ensuring that the model doesn’t conflate subtleties (e.g., pinching vs. scratching skin) without good feature separation. \n
\nThis notebook is a work in progress. I'm currently focusing on exploratory data analysis and will be extending it with preprocessing steps and model development soon.
\n📧 Email: tarunpmishra2001@gmail.com
\n\n🔗 LinkedIn: linkedin.com/in/tarunpmishra
\n\n🌐 Portfolio: tarundirector.github.io
\n\n