app.py

import torch
import gradio as gr
import pandas as pd
import plotly.graph_objects as go
from transformers import pipeline
import numpy as np
from tqdm.auto import tqdm
import warnings
import os
from datetime import datetime, timedelta
from scipy.stats import pearsonr
import ast 
warnings.simplefilter(action='ignore', category=FutureWarning)

DEVELOPER_NAME = "汪于捷、李哲弘、黃千宥、陳奕瑄、洪寓澤"

NEWS_CSV_PATH = 'cryptonews.csv'
BTC_CSV_PATH = 'BTC.csv'
PROCESSED_DATA_PATH = 'processed_btc_sentiment_data.csv'
PLOTLY_TEMPLATE = "plotly_dark"
SENTIMENT_PIPELINE = None

def initialize_pipeline():
    """載入情緒分析模型，只在需要時執行一次。"""
    global SENTIMENT_PIPELINE
    if SENTIMENT_PIPELINE is None:
        try:
            print("⏳ 正在載入情緒分析模型 (Hugging Face)...")
            MODEL_NAME = "cardiffnlp/twitter-roberta-base-sentiment"
            SENTIMENT_PIPELINE = pipeline(
                "sentiment-analysis", model=MODEL_NAME, tokenizer=MODEL_NAME, device=-1
            )
            print("✅ 模型載入成功！")
        except Exception as e:
            print(f"❌ 載入模型時發生錯誤: {e}")
            SENTIMENT_PIPELINE = None

def safe_literal_eval(val):
    """安全地解析字串，如果失敗則回傳空字典。"""
    try:
        return ast.literal_eval(val)
    except (ValueError, SyntaxError):
        return {}

def preprocess_and_cache_data():
    """
    執行一次性的資料預處理，分析來源為新聞標題(title)與內文(text)的組合。
    """
    if not os.path.exists(NEWS_CSV_PATH) or not os.path.exists(BTC_CSV_PATH):
        raise FileNotFoundError(f"請確認 '{NEWS_CSV_PATH}' 和 '{BTC_CSV_PATH}' 檔案存在。")

    initialize_pipeline()
    if SENTIMENT_PIPELINE is None:
        raise RuntimeError("情緒分析模型未能成功初始化。")

    print(f"⏳ 正在讀取原始資料: '{NEWS_CSV_PATH}'...")
    news_df = pd.read_csv(NEWS_CSV_PATH)
    news_df.dropna(subset=['title', 'text', 'sentiment'], inplace=True)
    news_df['date'] = pd.to_datetime(news_df['date'], errors='coerce').dt.date
    news_df.dropna(subset=['date'], inplace=True)

    print("⏳ 正在合併新聞標題與內文...")
    news_df['full_text'] = news_df['title'] + ". " + news_df['text']

    print("⏳ 正在對新聞完整內容 (標題+內文) 進行模型情緒分析...")
    texts_to_analyze = news_df['full_text'].tolist()
    sentiments_model = SENTIMENT_PIPELINE(
        texts_to_analyze, 
        batch_size=256, 
        truncation=True, 
        max_length=512
    )
    score_map_model = {'LABEL_2': 1, 'LABEL_1': 0, 'LABEL_0': -1}
    news_df['model_sentiment_score'] = [score_map_model.get(s['label'], 0) for s in sentiments_model]

    print("⏳ 正在解析預存的情緒欄位 (class, polarity, subjectivity)...")
    sentiment_dicts = news_df['sentiment'].apply(safe_literal_eval)
    
    class_score_map = {'positive': 1, 'neutral': 0, 'negative': -1}
    news_df['class_sentiment_score'] = sentiment_dicts.apply(lambda x: class_score_map.get(x.get('class', 'neutral'), 0))
    
    news_df['polarity'] = sentiment_dicts.apply(lambda x: x.get('polarity', 0.0))
    news_df['subjectivity'] = sentiment_dicts.apply(lambda x: x.get('subjectivity', 0.0))

    print("⏳ 正在計算每日平均情緒指標...")
    daily_metrics = news_df.groupby('date').agg(
        avg_model_sentiment=('model_sentiment_score', 'mean'),
        avg_class_sentiment=('class_sentiment_score', 'mean'),
        avg_polarity=('polarity', 'mean'),
        avg_subjectivity=('subjectivity', 'mean')
    ).reset_index()

    print(f"⏳ 正在讀取比特幣價格資料: '{BTC_CSV_PATH}'...")
    btc_df = pd.read_csv(BTC_CSV_PATH)
    btc_df['date'] = pd.to_datetime(btc_df['date'], errors='coerce').dt.date
    btc_df['price_change_pct'] = btc_df['close'].pct_change() * 100

    print("⏳ 正在合併所有資料...")
    daily_metrics['date'] = pd.to_datetime(daily_metrics['date'])
    btc_df['date'] = pd.to_datetime(btc_df['date'])
    merged_df = pd.merge(btc_df, daily_metrics, on='date', how='inner')
    
    news_content_df = news_df.groupby('date').agg(
        titles=('title', list),
        texts=('text', list)
    ).reset_index()
    news_content_df['date'] = pd.to_datetime(news_content_df['date'])
    
    final_df = pd.merge(merged_df, news_content_df, on='date', how='left')

    print(f"✅ 資料預處理完成！正在將結果儲存至 '{PROCESSED_DATA_PATH}'...")
    final_df.to_csv(PROCESSED_DATA_PATH, index=False)
    return final_df

def load_data():
    """載入資料，若快取不存在則執行預處理。"""
    if os.path.exists(PROCESSED_DATA_PATH):
        print(f"✅ 發現已處理的資料快取，正在從 '{PROCESSED_DATA_PATH}' 載入...")
        df = pd.read_csv(PROCESSED_DATA_PATH)
        df['date'] = pd.to_datetime(df['date'])
        df['titles'] = df['titles'].apply(ast.literal_eval)
        df['texts'] = df['texts'].apply(ast.literal_eval)
        return df
    else:
        print("⚠️ 未發現已處理的資料，將執行首次預處理...")
        return preprocess_and_cache_data()

df = load_data()
# 確保資料按日期排序
df.sort_values(by='date', inplace=True)
df.set_index('date', inplace=True)

def get_filtered_df(start_date, end_date):
    """根據日期範圍篩選 DataFrame。"""
    if start_date is None or end_date is None:
        return pd.DataFrame()
    return df[(df.index >= pd.to_datetime(start_date)) & (df.index <= pd.to_datetime(end_date))].copy()

def plot_price_and_sentiment(filtered_df, sentiment_col, sentiment_name, color):
    fig = go.Figure()
    fig.add_trace(go.Scatter(x=filtered_df.index, y=filtered_df['close'], name='BTC 收盤價', line=dict(color='deepskyblue'), yaxis='y1'))
    fig.add_trace(go.Scatter(x=filtered_df.index, y=filtered_df[sentiment_col], name=sentiment_name, line=dict(color=color, dash='dash'), yaxis='y2'))
    fig.update_layout(
        # title=f'📈 比特幣價格 vs. {sentiment_name}趨勢',
        xaxis_title='日期',
        yaxis=dict(title='價格 (USD)', color='deepskyblue'),
        yaxis2=dict(title='情緒分數', overlaying='y', side='right', color=color, range=[-1, 1]),
        legend=dict(x=0.01, y=0.99, orientation='h'),
        template=PLOTLY_TEMPLATE, 
        paper_bgcolor='rgba(0,0,0,0)', plot_bgcolor='rgba(0,0,0,0.2)'
    )
    return fig

def plot_subjectivity_trend(filtered_df):
    fig = go.Figure()
    fig.add_trace(go.Scatter(x=filtered_df.index, y=filtered_df['avg_subjectivity'], name='每日新聞主觀性', line=dict(color='lightgreen')))
    fig.update_layout(
        # title='🧐 每日新聞主觀性趨勢',
        xaxis_title='日期',
        yaxis=dict(title='主觀性分數 (0=客觀, 1=主觀)', color='lightgreen', range=[0, 1]),
        template=PLOTLY_TEMPLATE, 
        paper_bgcolor='rgba(0,0,0,0)', plot_bgcolor='rgba(0,0,0,0.2)'
    )
    return fig

def plot_correlation(filtered_df, sentiment_col, lag_days):
    df_corr = filtered_df[[sentiment_col, 'price_change_pct']].copy()
    df_corr['price_change_pct_lagged'] = df_corr['price_change_pct'].shift(-lag_days)
    df_corr.dropna(inplace=True)
    if df_corr.empty or len(df_corr) < 2:
        correlation, p_value = 0, 1
    else:
        correlation, p_value = pearsonr(df_corr[sentiment_col], df_corr['price_change_pct_lagged'])
    fig = go.Figure(data=go.Scatter(x=df_corr[sentiment_col], y=df_corr['price_change_pct_lagged'], mode='markers', marker=dict(color='mediumpurple', opacity=0.7)))
    fig.update_layout(
        title=f'🔗 情緒與 {lag_days} 天後價格變化的關聯性 (相關係數: {correlation:.3f})',
        xaxis_title='每日平均情緒分數', yaxis_title=f'{lag_days} 天後價格變化 (%)',
        template=PLOTLY_TEMPLATE, 
        paper_bgcolor='rgba(0,0,0,0)', plot_bgcolor='rgba(0,0,0,0.2)'
    )
    return fig, correlation, p_value

def get_top_bottom_news(date_obj):
    """
    獲取指定日期的最正面與最負面新聞。
    """
    date_ts = pd.to_datetime(date_obj)
    if date_ts not in df.index: 
        return "<ul><li>無此日期資料</li></ul>", "<ul><li>無此日期資料</li></ul>"
    
    day_data = df.loc[date_ts]
    titles, texts = day_data.get('titles', []), day_data.get('texts', [])
    
    initialize_pipeline()
    if SENTIMENT_PIPELINE is None or not isinstance(titles, list) or not isinstance(texts, list) or len(titles) != len(texts):
        return "<ul><li>模型未載入或新聞資料格式錯誤</li></ul>", "<ul><li>模型未載入或新聞資料格式錯誤</li></ul>"
        
    full_texts_for_day = [f"{title}. {text}" for title, text in zip(titles, texts)]
    if not full_texts_for_day: 
        return "<ul><li>當日無新聞</li></ul>", "<ul><li>當日無新聞</li></ul>"
        
    sentiments = SENTIMENT_PIPELINE(full_texts_for_day, batch_size=8, truncation=True, max_length=512)
    score_map = {'LABEL_2': 1, 'LABEL_1': 0, 'LABEL_0': -1}
    
    scored_titles = []
    for i, sentiment in enumerate(sentiments):
        directional_score = score_map.get(sentiment['label'], 0) * sentiment['score']
        scored_titles.append((titles[i], directional_score))
        
    positive_news = sorted([item for item in scored_titles if item[1] > 0], key=lambda x: x[1], reverse=True)
    negative_news = sorted([item for item in scored_titles if item[1] < 0], key=lambda x: x[1], reverse=False)

    if positive_news:
        top_news_html = "".join([f"<li>{title}</li>" for title, score in positive_news[:3]])
    else:
        top_news_html = "<li>當日無正面情緒新聞</li>"
        
    if negative_news:
        bottom_news_html = "".join([f"<li>{title}</li>" for title, score in negative_news[:3]])
    else:
        bottom_news_html = "<li>當日無負面情緒新聞</li>"
        
    return f"<ul>{top_news_html}</ul>", f"<ul>{bottom_news_html}</ul>"

with gr.Blocks(
    theme=gr.themes.Soft(
        primary_hue="sky", 
        secondary_hue="orange", 
        font=["Arial", "sans-serif"]
    ), 
    js="""
function refresh() {
    const url = new URL(window.location);
    if (url.searchParams.get('__theme') !== 'dark') {
        url.searchParams.set('__theme', 'dark');
        window.location.href = url.href;
    }
}
"""
) as app:
    gr.Markdown(f"""<div style='text-align: center; padding: 20px; color: white;'><h1 style='font-size: 3em; color: #00BFFF;'>📈 Crypto Pulse</h1><p style='font-size: 1.2em; color: #A9A9A9;'>比特幣新聞情緒與價格分析儀表板</p><p style='font-size: 0.9em; color: #888;'>Designed by: {DEVELOPER_NAME}</p></div>""")

    max_date_dt = df.index.max()
    # 確保資料數足夠
    if len(df) > 360:
        min_date_dt = df.index[-360]
    else:
        min_date_dt = df.index.min()
    
    with gr.Row():
        start_date_input = gr.DateTime(label="📅 開始日期", type="datetime", value=min_date_dt)
        end_date_input = gr.DateTime(label="📅 結束日期", type="datetime", value=max_date_dt)
    
    with gr.Tabs() as tabs:
        with gr.TabItem("📊 模型情緒總覽", id=0):
            plot_overview = gr.Plot(label="模型情緒 vs. 價格趨勢圖")
            gr.Markdown("此圖展示了由 `twitter-roberta-base-sentiment` 模型分析出的**新聞內容（標題+內文）**情緒分數（右軸）與比特幣價格（左軸）的對比。")

        with gr.TabItem("🔬 多維度情緒分析", id=1):
            gr.Markdown("""
            ### 指標說明
            此處的情緒指標來自資料集 `cryptonews.csv` 中預先計算好的 `sentiment` 欄位。
            * **資料集預設情緒分類**: 將資料集內建的 `positive`, `neutral`, `negative` 類別轉換為 `1, 0, -1` 的數值分數。
            * **情感極性 (Polarity)**: 衡量文本的正面或負面程度。值域從 -1 (非常負面) 到 +1 (非常正面)。
            * **主觀性 (Subjectivity)**: 衡量文本是偏向客觀事實還是主觀意見。值域從 0 (非常客觀) 到 1 (非常主觀)。
            """)
            plot_class_sentiment = gr.Plot(label="資料集預設情緒 vs. 價格趨勢圖")
            plot_polarity = gr.Plot(label="情感極性 vs. 價格趨勢圖")
            plot_subjectivity = gr.Plot(label="新聞主觀性趨勢圖")

        with gr.TabItem("🔍 關聯性深掘", id=2):
            with gr.Row():
                with gr.Column(scale=2, min_width=250):
                    sentiment_type_radio = gr.Radio(
                        ["模型情緒分數", "資料集預設情緒分類", "情感極性 (Polarity)"],
                        label="選擇分析的情緒指標", value="模型情緒分數"
                    )
                    lag_slider = gr.Slider(minimum=0, maximum=14, value=1, step=1, label="🕒 情緒延遲天數 (Lag Days)")
                    correlation_output = gr.Textbox(label="Pearson 相關係數", interactive=False)
                    p_value_output = gr.Textbox(label="P-Value", interactive=False)
                with gr.Column(scale=3):
                    plot_corr = gr.Plot(label="情緒 vs. 價格變化 散點圖")
        
        with gr.TabItem("📰 新聞瀏覽器", id=3):
            gr.Markdown("在此處選擇特定日期，即可查看當天的熱點新聞。")
            news_date_input = gr.DateTime(label="🗓️ 選擇查詢日期", type="datetime", value=max_date_dt)
            with gr.Row():
                gr.Markdown("### 👍 當日最正面新聞 Top 3"); gr.Markdown("### 👎 當日最負面新聞 Top 3")
            with gr.Row():
                top_news_output = gr.HTML(); bottom_news_output = gr.HTML()

    def update_all(start_date, end_date, lag_days, sentiment_type):
        if start_date is None or end_date is None or start_date > end_date:
            gr.Warning("請選擇有效的開始與結束日期。")
            empty_fig = go.Figure()
            return empty_fig, empty_fig, empty_fig, empty_fig, empty_fig, "N/A", "N/A"

        start_date, end_date = pd.to_datetime(start_date), pd.to_datetime(end_date)
        filtered_df = get_filtered_df(start_date, end_date)
        if filtered_df.empty:
            gr.Warning("此日期範圍內無資料，請擴大範圍。")
            empty_fig = go.Figure()
            return empty_fig, empty_fig, empty_fig, empty_fig, empty_fig, "N/A", "N/A"
        
        overview_fig = plot_price_and_sentiment(filtered_df, 'avg_model_sentiment', '模型情緒分數', 'crimson')
        class_sentiment_fig = plot_price_and_sentiment(filtered_df, 'avg_class_sentiment', '資料集預設情緒分類', 'yellow')
        polarity_fig = plot_price_and_sentiment(filtered_df, 'avg_polarity', '情感極性 (Polarity)', 'orange')
        subjectivity_fig = plot_subjectivity_trend(filtered_df)
        
        if sentiment_type == "模型情緒分數":
            sentiment_col = 'avg_model_sentiment'
        elif sentiment_type == "資料集預設情緒分類":
            sentiment_col = 'avg_class_sentiment'
        else: # Polarity
            sentiment_col = 'avg_polarity'
            
        corr_fig, corr_val, p_val = plot_correlation(filtered_df, sentiment_col, lag_days)
        
        return overview_fig, class_sentiment_fig, polarity_fig, subjectivity_fig, corr_fig, f"{corr_val:.4f}", f"{p_val:.4f}"

    def update_news_browser(date_obj):
        if date_obj is None:
            return "請選擇日期", "無"
        top_news, bottom_news = get_top_bottom_news(date_obj)
        return top_news, bottom_news

    inputs_for_main_update = [start_date_input, end_date_input, lag_slider, sentiment_type_radio]
    outputs_for_main_update = [plot_overview, plot_class_sentiment, plot_polarity, plot_subjectivity, plot_corr, correlation_output, p_value_output]
    
    for component in [start_date_input, end_date_input, lag_slider, sentiment_type_radio]:
        component.change(fn=update_all, inputs=inputs_for_main_update, outputs=outputs_for_main_update)
    
    news_date_input.change(
        fn=update_news_browser,
        inputs=[news_date_input],
        outputs=[top_news_output, bottom_news_output]
    )

    def load_app():
        main_outputs = update_all(min_date_dt, max_date_dt, 1, "模型情緒分數")
        news_outputs = update_news_browser(max_date_dt)
        return main_outputs + news_outputs

    app.load(
        fn=load_app, 
        inputs=None, 
        outputs=outputs_for_main_update + [top_news_output, bottom_news_output]
    )

app.launch(debug=False, share=True, show_error=True, show_api=False)