import streamlit as st
import pandas as pd
import numpy as np
import altair as alt
import plotly.express as px
from datasets import load_dataset
from scipy.stats import gaussian_kde

st.set_page_config(
    page_title="Mobitag Explorer",
    page_icon="📱",
    layout="wide",
)

st.title("📱 Mobitag — Activité SMS Nouvelle-Calédonie")
st.caption("Dataset : [opt-nc/mobitag](https://huggingface.co/datasets/opt-nc/mobitag) · Heure locale NC (UTC+11)")


@st.cache_data(show_spinner="Chargement du dataset…")
def load_data():
    ds = load_dataset("opt-nc/mobitag")
    df = pd.concat([ds["train"].to_pandas(), ds["test"].to_pandas()])
    df["date"] = pd.to_datetime(df["date"])
    df["date_local"] = df["date"] + pd.Timedelta(hours=11)
    return df


df = load_data()

day_fr = {"Monday":"Lundi","Tuesday":"Mardi","Wednesday":"Mercredi",
          "Thursday":"Jeudi","Friday":"Vendredi","Saturday":"Samedi","Sunday":"Dimanche"}
day_order_fr = ["Lundi","Mardi","Mercredi","Jeudi","Vendredi","Samedi","Dimanche"]
month_names = {1:"Jan",2:"Fév",3:"Mar",4:"Avr",5:"Mai",6:"Jun",
               7:"Jul",8:"Aoû",9:"Sep",10:"Oct",11:"Nov",12:"Déc"}

# --- Sidebar filtres ---
st.sidebar.header("Filtres")
years = sorted(df["year"].unique())
selected_years = st.sidebar.multiselect("Année", years, default=years)
df_f = df[df["year"].isin(selected_years)]

st.sidebar.markdown(f"**{len(df_f):,} événements** sélectionnés")

# --- KPIs ---
col1, col2, col3, col4 = st.columns(4)
col1.metric("Total SMS", f"{len(df_f):,}")
col2.metric("Longueur moyenne", f"{df_f['message_length'].mean():.0f} car.")
col3.metric("Longueur médiane", f"{df_f['message_length'].median():.0f} car.")
col4.metric("Période couverte", f"{df_f['year_month'].min()} → {df_f['year_month'].max()}")

st.divider()

# --- Volume journalier ---
st.subheader("Volume journalier de SMS")
daily_vol = df_f.groupby(df_f["date_local"].dt.date).size().reset_index(name="count")
daily_vol.columns = ["date", "count"]
daily_vol["date"] = pd.to_datetime(daily_vol["date"])
chart_daily = alt.Chart(daily_vol).mark_line(strokeWidth=1, opacity=0.8, color="#1f77b4").encode(
    x=alt.X("date:T", title="Date"),
    y=alt.Y("count:Q", title="SMS / jour"),
    tooltip=[alt.Tooltip("date:T", title="Date", format="%d/%m/%Y"), alt.Tooltip("count:Q", title="SMS")],
).properties(height=280)
st.altair_chart(chart_daily, use_container_width=True)

st.divider()

# --- Volume mensuel ---
st.subheader("Volume mensuel de SMS")
monthly = df_f.groupby("year_month").size().reset_index(name="count")
chart = alt.Chart(monthly).mark_bar().encode(
    x=alt.X("year_month:O", title="Mois", axis=alt.Axis(labelAngle=-45)),
    y=alt.Y("count:Q", title="Nombre de SMS"),
    tooltip=["year_month", "count"],
).properties(height=300)
st.altair_chart(chart, use_container_width=True)

st.divider()

# --- Longueur moyenne des messages dans le temps ---
st.subheader("Longueur moyenne des messages par mois")
avg_length = df_f.groupby("year_month")["message_length"].mean().reset_index(name="avg")
avg_length["avg"] = avg_length["avg"].round(1)
chart_avg = alt.Chart(avg_length).mark_line(point=True, color="#2ca02c", strokeWidth=2).encode(
    x=alt.X("year_month:O", title="Mois", axis=alt.Axis(labelAngle=-45)),
    y=alt.Y("avg:Q", title="Longueur moyenne (car.)", scale=alt.Scale(zero=False)),
    tooltip=["year_month", alt.Tooltip("avg:Q", title="Longueur moy.", format=".1f")],
).properties(height=280)
st.altair_chart(chart_avg, use_container_width=True)

st.divider()

# --- Comparaison année par année (volume par mois) ---
st.subheader("Comparaison année par année — volume mensuel")

# Détecte l'année incomplète (dernier mois < décembre)
max_year = int(df["year"].max())
max_month_of_last_year = int(df[df["year"] == max_year]["month"].max())
if max_month_of_last_year < 12:
    st.info(f"ℹ️ **{max_year} est une année incomplète** — données disponibles jusqu'en {month_names[max_month_of_last_year]} {max_year} seulement.")

yoy = df_f.groupby(["year", "month"]).size().reset_index(name="count")
yoy["year"] = yoy["year"].astype(str)
yoy["mois"] = yoy["month"].map(month_names)
chart_yoy = alt.Chart(yoy).mark_bar().encode(
    x=alt.X("month:O", title="Mois", axis=alt.Axis(labelExpr=
        "['Jan','Fév','Mar','Avr','Mai','Jun','Jul','Aoû','Sep','Oct','Nov','Déc'][datum.value-1]"
    )),
    xOffset=alt.XOffset("year:N"),
    y=alt.Y("count:Q", title="Nombre de SMS"),
    color=alt.Color("year:N", title="Année", scale=alt.Scale(scheme="tableau10")),
    tooltip=["year", "mois", "count"],
).properties(height=320)
st.altair_chart(chart_yoy, use_container_width=True)

st.divider()

# --- Distribution horaire & jour de semaine ---
col_l, col_r = st.columns(2)

with col_l:
    st.subheader("Activité par heure (local NC)")
    hourly = df_f.groupby("hour").agg(count=("message_length","count"), avg_len=("message_length","mean")).reset_index()
    hourly["avg_len"] = hourly["avg_len"].round(1)
    chart_h = alt.Chart(hourly).mark_bar(color="#1f77b4").encode(
        x=alt.X("hour:O", title="Heure"),
        y=alt.Y("count:Q", title="Nombre de SMS"),
        tooltip=["hour", "count", alt.Tooltip("avg_len:Q", title="Longueur moy.", format=".1f")],
    ).properties(height=200)
    base = alt.Chart(hourly).encode(
        x=alt.X("hour:Q", title="Heure"),
        y=alt.Y("avg_len:Q", title="Longueur moy. (car.)", scale=alt.Scale(zero=False)),
    )
    area = base.mark_area(color="#2ca02c", opacity=0.15)
    line = base.mark_line(color="#2ca02c", strokeWidth=2)
    trend = base.transform_loess("hour", "avg_len", bandwidth=0.4).mark_line(
        color="#d62728", strokeWidth=2, strokeDash=[5, 3]
    )
    chart_len_h = (area + line + trend).properties(height=160)
    st.altair_chart(chart_h & chart_len_h, use_container_width=True)

with col_r:
    st.subheader("Activité par jour de semaine")
    daily = df_f.groupby("day_name").size().reset_index(name="count")
    daily["jour"] = daily["day_name"].map(day_fr)
    chart_d = alt.Chart(daily).mark_bar(color="#ff7f0e").encode(
        x=alt.X("jour:O", title="Jour", sort=day_order_fr),
        y=alt.Y("count:Q", title="Nombre de SMS"),
        tooltip=["jour", "count"],
    ).properties(height=280)
    st.altair_chart(chart_d, use_container_width=True)

st.divider()

# --- Distribution longueur des messages ---
st.subheader("Distribution de la longueur des messages")
hist = alt.Chart(df_f.sample(min(50_000, len(df_f)))).mark_bar(opacity=0.7).encode(
    x=alt.X("message_length:Q", bin=alt.Bin(maxbins=40), title="Longueur (caractères)"),
    y=alt.Y("count():Q", title="Fréquence"),
    tooltip=["count()"],
).properties(height=280)
st.altair_chart(hist, use_container_width=True)

st.divider()

# --- Sunburst jour × heure ---
st.subheader("Sunburst : distribution jour de semaine / heure")
sb_data = df_f.groupby(["day_name", "hour"]).size().reset_index(name="count")
sb_data["jour"] = pd.Categorical(sb_data["day_name"].map(day_fr), categories=day_order_fr, ordered=True)
sb_data["heure"] = sb_data["hour"].astype(str).str.zfill(2) + "h"
sb_data = sb_data.sort_values(["jour", "hour"])
fig = px.sunburst(
    sb_data,
    path=["jour", "heure"],
    values="count",
    color="count",
    color_continuous_scale="Blues",
)
fig.update_layout(margin=dict(t=10, b=10, l=10, r=10), height=550)
st.plotly_chart(fig, use_container_width=True)

st.divider()

# --- Heatmap heure × jour ---
st.subheader("Heatmap : heure × jour de semaine")
heatmap_data = df_f.groupby(["day_name", "hour"]).size().reset_index(name="count")
heatmap_data["jour"] = heatmap_data["day_name"].map(day_fr)
heatmap = alt.Chart(heatmap_data).mark_rect().encode(
    x=alt.X("hour:O", title="Heure (local NC)"),
    y=alt.Y("jour:O", title="Jour", sort=day_order_fr),
    color=alt.Color("count:Q", scale=alt.Scale(scheme="blues"), title="SMS"),
    tooltip=["jour", "hour", "count"],
).properties(height=250)
st.altair_chart(heatmap, use_container_width=True)

st.divider()

# --- Détermination du seuil optimal (Jenks + KMeans) ---
st.subheader("📐 Détermination du seuil optimal — Jenks & KMeans")
st.caption("Les deux méthodes convergent à **99 caractères** : frontière naturelle entre deux populations distinctes.")

SPLIT = 99
CENTER_SHORT, CENTER_LONG = 61.5, 137.0

col_a, col_b = st.columns(2)
col_a.metric("Centroïde messages courts", f"{CENTER_SHORT:.0f} chars", "cluster 1")
col_b.metric("Centroïde messages longs", f"{CENTER_LONG:.0f} chars", "cluster 2")

lengths = df["message_length"]

# Double density plot — KDE pré-calculée côté serveur (scipy)
x_range = np.linspace(1, 164, 300)
kde_short = gaussian_kde(lengths[lengths <= SPLIT].values, bw_method=0.15)
kde_long  = gaussian_kde(lengths[lengths > SPLIT].values,  bw_method=0.15)

kde_df = pd.DataFrame({
    "length":  np.concatenate([x_range, x_range]),
    "density": np.concatenate([kde_short(x_range), kde_long(x_range)]),
    "cluster": ["Non authentifié (≤ 99)"] * 300 + ["Authentifié (> 99)"] * 300,
})

color_scale = alt.Scale(
    domain=["Non authentifié (≤ 99)", "Authentifié (> 99)"],
    range=["#1f77b4", "#d62728"]
)
area_density = alt.Chart(kde_df).mark_area(opacity=0.35).encode(
    x=alt.X("length:Q", title="Longueur (caractères)"),
    y=alt.Y("density:Q", title="Densité", stack=None),
    color=alt.Color("cluster:N", scale=color_scale, title="Cluster"),
)
line_density = alt.Chart(kde_df).mark_line(strokeWidth=2).encode(
    x="length:Q",
    y=alt.Y("density:Q", stack=None),
    color=alt.Color("cluster:N", scale=color_scale, legend=None),
)
centroid_lines = alt.Chart(
    pd.DataFrame({"x": [CENTER_SHORT, CENTER_LONG],
                  "cluster": ["Non authentifié (≤ 99)", "Authentifié (> 99)"]})
).mark_rule(strokeDash=[5, 3], strokeWidth=1.5).encode(
    x="x:Q",
    color=alt.Color("cluster:N", scale=color_scale, legend=None),
)
split_rule = alt.Chart(pd.DataFrame({"x": [SPLIT]})).mark_rule(
    color="black", strokeWidth=2
).encode(x="x:Q")

density_chart = (area_density + line_density + centroid_lines + split_rule).properties(height=320)
st.altair_chart(density_chart, use_container_width=True)
st.caption("**Zone bleue** = cluster non authentifié (centroïde 62 chars) · **zone rouge** = cluster authentifié (centroïde 137 chars) · **ligne noire** = seuil à 99 chars · le chevauchement montre la frontière naturelle entre les deux populations")

st.divider()

# --- Analyse tarifaire freemium ---
st.subheader("💰 Analyse tarifaire — Split non authentifié / authentifié")
st.caption("Modèle : tier non authentifié (≤ seuil) monétisé par pub · tier authentifié (> seuil) sans pub")

AVG_MSGS_PER_MONTH = 63_100  # moyenne historique du dataset

col_sl1, col_sl2 = st.columns(2)
with col_sl1:
    threshold = st.slider("Seuil non authentifié / authentifié (caractères)", min_value=50, max_value=160,
                          value=99, step=1)
with col_sl2:
    XPF_PER_AD = st.slider("Revenu par affichage pub (XPF)", min_value=0.1, max_value=1.0,
                            value=1.0, step=0.1, format="%.1f XPF")

n_free    = int((lengths <= threshold).sum())
n_paid    = int((lengths > threshold).sum())
pct_free  = n_free / len(lengths) * 100
pct_paid  = n_paid / len(lengths) * 100

msgs_free_monthly = AVG_MSGS_PER_MONTH * pct_free / 100
revenue_monthly   = msgs_free_monthly * XPF_PER_AD
revenue_annual    = revenue_monthly * 12

c1, c2, c3, c4 = st.columns(4)
c1.metric("Messages non authentifiés (pub)", f"{pct_free:.1f}%", f"{msgs_free_monthly:,.0f} msgs/mois")
c2.metric("Messages authentifiés", f"{pct_paid:.1f}%", f"{AVG_MSGS_PER_MONTH - msgs_free_monthly:,.0f} msgs/mois")
c3.metric("💰 Gain pub / mois", f"{revenue_monthly:,.0f} XPF", f"≈ {revenue_monthly/119.33:.0f} €")
c4.metric("💰 Gain pub / an", f"{revenue_annual:,.0f} XPF", f"≈ {revenue_annual/119.33:.0f} €")

# Courbe CDF avec seuil
cdf_data = pd.DataFrame({
    "length": range(1, 165),
    "pct_free": [(lengths <= t).sum() / len(lengths) * 100 for t in range(1, 165)],
})
rule = alt.Chart(pd.DataFrame({"x": [threshold]})).mark_rule(
    color="red", strokeWidth=2, strokeDash=[6, 3]
).encode(x="x:Q")
cdf_line = alt.Chart(cdf_data).mark_line(color="#1f77b4", strokeWidth=2).encode(
    x=alt.X("length:Q", title="Longueur message (caractères)"),
    y=alt.Y("pct_free:Q", title="% messages non authentifiés (CDF)", scale=alt.Scale(domain=[0, 100])),
    tooltip=["length", alt.Tooltip("pct_free:Q", format=".1f", title="% non authentifié")],
)
st.altair_chart(cdf_line + rule, use_container_width=True)
st.caption(f"La ligne rouge marque le seuil à **{threshold} caractères** — {pct_free:.1f}% des messages sont non authentifiés, {pct_paid:.1f}% sont authentifiés.")