app.py

"""
🔔 Notification Bad-Timing Detector — Interactive Demo
Predicts the probability that NOW is a bad time to send a push notification.
Uses a LightGBM model trained on 100K samples with 21 contextual features.
"""

import math
import pickle
import numpy as np
import pandas as pd
import gradio as gr
from huggingface_hub import hf_hub_download

# ── Load model from HF Hub ──────────────────────────────────────────────────
MODEL_REPO = "alianassmaaa/notification-bad-timing-detector"
model_path = hf_hub_download(repo_id=MODEL_REPO, filename="calibrated_model.pkl")
with open(model_path, "rb") as f:
    model = pickle.load(f)

# ── Feature order (must match training) ──────────────────────────────────────
FEATURES = [
    "hour_of_day", "day_of_week", "hour_sin", "hour_cos",
    "is_weekend", "is_night",
    "battery_level", "is_charging", "battery_change_rate",
    "screen_on", "screen_on_duration_30min", "app_opens_last_hour",
    "session_length_current", "time_since_last_interaction",
    "notif_shown_last_30min", "notif_clicked_last_30min",
    "notif_dismissed_last_30min", "notif_ignored_last_30min",
    "notif_shown_last_24h", "notif_ctr_last_7d",
    "recent_notification_density",
]

DAY_NAMES = ["Monday", "Tuesday", "Wednesday", "Thursday", "Friday", "Saturday", "Sunday"]

# ── Prediction function ──────────────────────────────────────────────────────
def predict(
    hour_of_day, day_of_week,
    battery_level, is_charging, battery_change_rate,
    screen_on, screen_on_duration_30min, app_opens_last_hour,
    session_length_current, time_since_last_interaction,
    notif_shown_last_30min, notif_clicked_last_30min,
    notif_dismissed_last_30min, notif_ignored_last_30min,
    notif_shown_last_24h, notif_ctr_last_7d,
    recent_notification_density,
):
    # Derived features
    day_idx = DAY_NAMES.index(day_of_week)
    hour_sin = math.sin(2 * math.pi * hour_of_day / 24)
    hour_cos = math.cos(2 * math.pi * hour_of_day / 24)
    is_weekend = 1 if day_idx >= 5 else 0
    is_night = 1 if (hour_of_day >= 22 or hour_of_day < 7) else 0

    features = pd.DataFrame([[
        hour_of_day, day_idx, hour_sin, hour_cos,
        is_weekend, is_night,
        battery_level, int(is_charging), battery_change_rate,
        int(screen_on), screen_on_duration_30min, app_opens_last_hour,
        session_length_current, time_since_last_interaction,
        notif_shown_last_30min, notif_clicked_last_30min,
        notif_dismissed_last_30min, notif_ignored_last_30min,
        notif_shown_last_24h, notif_ctr_last_7d,
        recent_notification_density,
    ]], columns=FEATURES)

    prob = model.predict_proba(features)[:, 1][0]

    # Build result
    if prob < 0.3:
        verdict = "✅ Good time to send!"
        advice = "The user is likely receptive. Send the notification now."
    elif prob < 0.5:
        verdict = "⚠️ Consider priority"
        advice = "Send only if the notification is important or time-sensitive."
    elif prob < 0.8:
        verdict = "🚫 Bad timing — delay"
        advice = "The user is likely busy or disengaged. Schedule for later."
    else:
        verdict = "🔴 Definitely delay!"
        advice = "Very bad timing. The notification will almost certainly be ignored or dismissed."

    # Gauge-like HTML output
    bar_width = int(prob * 100)
    bar_color = f"hsl({int((1 - prob) * 120)}, 80%, 45%)"

    html = f"""
    <div style="font-family: sans-serif; max-width: 500px; margin: auto;">
        <div style="text-align: center; margin-bottom: 16px;">
            <span style="font-size: 3em; font-weight: bold; color: {bar_color};">{prob:.1%}</span>
            <br/>
            <span style="font-size: 1.3em; font-weight: 600;">{verdict}</span>
        </div>
        <div style="background: #e0e0e0; border-radius: 12px; height: 28px; overflow: hidden; margin-bottom: 12px;">
            <div style="background: {bar_color}; width: {bar_width}%; height: 100%; border-radius: 12px; transition: width 0.4s;"></div>
        </div>
        <div style="display: flex; justify-content: space-between; font-size: 0.85em; color: #666; margin-bottom: 16px;">
            <span>0% — Great time</span>
            <span>100% — Terrible time</span>
        </div>
        <div style="background: #f8f9fa; border-left: 4px solid {bar_color}; padding: 12px 16px; border-radius: 4px;">
            <strong>Recommendation:</strong> {advice}
        </div>
    </div>
    """
    return html


# ── Preset scenarios ─────────────────────────────────────────────────────────
PRESETS = {
    "🌅 Morning commute (good time)": {
        "hour_of_day": 8, "day_of_week": "Tuesday",
        "battery_level": 90, "is_charging": False, "battery_change_rate": -0.5,
        "screen_on": True, "screen_on_duration_30min": 600, "app_opens_last_hour": 5,
        "session_length_current": 120, "time_since_last_interaction": 10,
        "notif_shown_last_30min": 1, "notif_clicked_last_30min": 1,
        "notif_dismissed_last_30min": 0, "notif_ignored_last_30min": 0,
        "notif_shown_last_24h": 15, "notif_ctr_last_7d": 0.45,
        "recent_notification_density": 0.5,
    },
    "😴 Late night sleeping (bad time)": {
        "hour_of_day": 3, "day_of_week": "Wednesday",
        "battery_level": 65, "is_charging": True, "battery_change_rate": 1.5,
        "screen_on": False, "screen_on_duration_30min": 0, "app_opens_last_hour": 0,
        "session_length_current": 0, "time_since_last_interaction": 7200,
        "notif_shown_last_30min": 0, "notif_clicked_last_30min": 0,
        "notif_dismissed_last_30min": 0, "notif_ignored_last_30min": 0,
        "notif_shown_last_24h": 20, "notif_ctr_last_7d": 0.15,
        "recent_notification_density": 0.0,
    },
    "📱 Active browsing (good time)": {
        "hour_of_day": 14, "day_of_week": "Saturday",
        "battery_level": 72, "is_charging": False, "battery_change_rate": -1.2,
        "screen_on": True, "screen_on_duration_30min": 1200, "app_opens_last_hour": 8,
        "session_length_current": 300, "time_since_last_interaction": 5,
        "notif_shown_last_30min": 2, "notif_clicked_last_30min": 2,
        "notif_dismissed_last_30min": 0, "notif_ignored_last_30min": 0,
        "notif_shown_last_24h": 18, "notif_ctr_last_7d": 0.52,
        "recent_notification_density": 1.0,
    },
    "🔋 Low battery, ignoring notifs (bad time)": {
        "hour_of_day": 19, "day_of_week": "Friday",
        "battery_level": 8, "is_charging": False, "battery_change_rate": -3.0,
        "screen_on": False, "screen_on_duration_30min": 60, "app_opens_last_hour": 1,
        "session_length_current": 0, "time_since_last_interaction": 1800,
        "notif_shown_last_30min": 5, "notif_clicked_last_30min": 0,
        "notif_dismissed_last_30min": 3, "notif_ignored_last_30min": 2,
        "notif_shown_last_24h": 45, "notif_ctr_last_7d": 0.08,
        "recent_notification_density": 4.0,
    },
    "🏢 Work meeting (bad time)": {
        "hour_of_day": 10, "day_of_week": "Monday",
        "battery_level": 55, "is_charging": False, "battery_change_rate": -0.8,
        "screen_on": False, "screen_on_duration_30min": 30, "app_opens_last_hour": 0,
        "session_length_current": 0, "time_since_last_interaction": 3600,
        "notif_shown_last_30min": 3, "notif_clicked_last_30min": 0,
        "notif_dismissed_last_30min": 1, "notif_ignored_last_30min": 2,
        "notif_shown_last_24h": 30, "notif_ctr_last_7d": 0.12,
        "recent_notification_density": 2.5,
    },
}

def load_preset(preset_name):
    if not preset_name or preset_name not in PRESETS:
        return [gr.update()] * 17
    p = PRESETS[preset_name]
    return [
        p["hour_of_day"], p["day_of_week"],
        p["battery_level"], p["is_charging"], p["battery_change_rate"],
        p["screen_on"], p["screen_on_duration_30min"], p["app_opens_last_hour"],
        p["session_length_current"], p["time_since_last_interaction"],
        p["notif_shown_last_30min"], p["notif_clicked_last_30min"],
        p["notif_dismissed_last_30min"], p["notif_ignored_last_30min"],
        p["notif_shown_last_24h"], p["notif_ctr_last_7d"],
        p["recent_notification_density"],
    ]


# ── Gradio UI ────────────────────────────────────────────────────────────────
with gr.Blocks(
    title="🔔 Notification Timing Detector",
    theme=gr.themes.Soft(),
) as demo:
    gr.Markdown("""
    # 🔔 Notification Bad-Timing Detector
    **Should you send that push notification right now?** This model predicts the probability that
    the current moment is a **bad time** to interrupt the user, based on their activity patterns,
    battery status, and notification interaction history.

    Built with LightGBM + isotonic calibration • Trained on 100K samples •
    [Model](https://huggingface.co/alianassmaaa/notification-bad-timing-detector) •
    [Dataset](https://huggingface.co/datasets/alianassmaaa/notification-timing-dataset)
    """)

    with gr.Row():
        preset = gr.Dropdown(
            choices=list(PRESETS.keys()),
            label="⚡ Quick presets — try a scenario",
            value=None,
            interactive=True,
        )

    with gr.Row():
        # Left column: inputs
        with gr.Column(scale=3):
            gr.Markdown("### ⏰ Time Context")
            with gr.Row():
                hour_of_day = gr.Slider(0, 23, value=14, step=1, label="Hour of day (0–23)")
                day_of_week = gr.Dropdown(DAY_NAMES, value="Wednesday", label="Day of week")

            gr.Markdown("### 🔋 Battery Status")
            with gr.Row():
                battery_level = gr.Slider(0, 100, value=75, step=1, label="Battery level (%)")
                is_charging = gr.Checkbox(value=False, label="Charging?")
                battery_change_rate = gr.Slider(-5, 5, value=-1.0, step=0.1, label="Battery drain rate (%/hr)")

            gr.Markdown("### 📱 User Activity")
            with gr.Row():
                screen_on = gr.Checkbox(value=True, label="Screen on?")
                screen_on_duration_30min = gr.Slider(0, 1800, value=600, step=10, label="Screen-on time last 30 min (sec)")
            with gr.Row():
                app_opens_last_hour = gr.Slider(0, 30, value=5, step=1, label="App opens last hour")
                session_length_current = gr.Slider(0, 3600, value=120, step=10, label="Current session length (sec)")
            time_since_last_interaction = gr.Slider(0, 14400, value=30, step=10, label="Time since last interaction (sec)")

            gr.Markdown("### 🔔 Notification History")
            with gr.Row():
                notif_shown_last_30min = gr.Slider(0, 20, value=2, step=1, label="Shown (30 min)")
                notif_clicked_last_30min = gr.Slider(0, 20, value=1, step=1, label="Clicked (30 min)")
            with gr.Row():
                notif_dismissed_last_30min = gr.Slider(0, 20, value=0, step=1, label="Dismissed (30 min)")
                notif_ignored_last_30min = gr.Slider(0, 20, value=1, step=1, label="Ignored (30 min)")
            with gr.Row():
                notif_shown_last_24h = gr.Slider(0, 100, value=20, step=1, label="Shown (24h)")
                notif_ctr_last_7d = gr.Slider(0, 1, value=0.35, step=0.01, label="7-day CTR")
            recent_notification_density = gr.Slider(0, 10, value=1.0, step=0.1, label="Recent notification density")

        # Right column: output
        with gr.Column(scale=2):
            predict_btn = gr.Button("🔮 Predict Bad-Timing Probability", variant="primary", size="lg")
            output = gr.HTML(label="Prediction")

            gr.Markdown("""
            ### 📊 Decision Thresholds
            | Probability | Action |
            |:-----------:|--------|
            | **< 30%** | ✅ Send notification |
            | **30–50%** | ⚠️ Only if important |
            | **50–80%** | 🚫 Delay notification |
            | **> 80%** | 🔴 Definitely delay |

            ---
            ### 🧠 How it works
            The model uses **21 signals** across 4 categories:
            - **Time**: hour, day, weekend/night flags
            - **Battery**: level, charging, drain rate
            - **Activity**: screen, app opens, session length
            - **Notifications**: recent shown/clicked/dismissed/ignored, 7-day CTR

            Built on research from the [C-3PO paper](https://arxiv.org/abs/1803.00458)
            (Cheetah Mobile, 600M monthly active users).
            """)

    # All input components in order
    all_inputs = [
        hour_of_day, day_of_week,
        battery_level, is_charging, battery_change_rate,
        screen_on, screen_on_duration_30min, app_opens_last_hour,
        session_length_current, time_since_last_interaction,
        notif_shown_last_30min, notif_clicked_last_30min,
        notif_dismissed_last_30min, notif_ignored_last_30min,
        notif_shown_last_24h, notif_ctr_last_7d,
        recent_notification_density,
    ]

    # Wire up events
    predict_btn.click(fn=predict, inputs=all_inputs, outputs=output)
    preset.change(fn=load_preset, inputs=preset, outputs=all_inputs)


if __name__ == "__main__":
    demo.launch()