prophet outliers异常值

Myth丶恋晨 2021-11-05 03:34 248阅读 0赞

# 例子代码 #

[https://github.com/lilihongjava/prophet\_demo/tree/master/outliers][https_github.com_lilihongjava_prophet_demo_tree_master_outliers]

# encoding: utf-8
    """
    @author: lee
    @time: 2019/8/8 9:35
    @file: main.py
    @desc: 
    """
    from fbprophet import Prophet
    import pandas as pd
    
    
    def main():
        df = pd.read_csv('./data/example_wp_log_R_outliers1.csv')
        m = Prophet()
        m.fit(df)
        future = m.make_future_dataframe(periods=1096)
        forecast = m.predict(future)
        fig = m.plot(forecast)
        fig.show()
    
        df.loc[(df['ds'] > '2010-01-01') & (df['ds'] < '2011-01-01'), 'y'] = None
        model = Prophet().fit(df)
        fig = model.plot(model.predict(future))
        fig.show()
    
        df = pd.read_csv('./data/example_wp_log_R_outliers2.csv')
        m = Prophet()
        m.fit(df)
        future = m.make_future_dataframe(periods=1096)
        forecast = m.predict(future)
        fig = m.plot(forecast)
        fig.show()
    
        df.loc[(df['ds'] > '2015-06-01') & (df['ds'] < '2015-06-30'), 'y'] = None
        m = Prophet().fit(df)
        fig = m.plot(m.predict(future))
        fig.show()
    
    
    if __name__ == "__main__":
        main()

异常值有两种主要方式可以影响prophet的预测。以下，我们对之前例子中记录维基百科中R语言页面访问数的数据进行预测，但是我们造了一段问题数据(2010-01到2010-05都是5，2010-06到2011-01都是8，2010-06-01为6.76272950693188)：

df = pd.read_csv('./data/example_wp_log_R_outliers1.csv')
    m = Prophet()
    m.fit(df)
    future = m.make_future_dataframe(periods=1096)
    forecast = m.predict(future)
    fig = m.plot(forecast)

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzMzODczNDMx_size_16_color_FFFFFF_t_70][]

趋势预测看似合理，但不确定性区间似乎过于宽泛。Prophet能够处理历史数据中的异常值，但只能通过趋势变化来拟合它们。不确定性模型预计未来趋势变化的幅度与历史是相似的。

处理异常值的最佳方法是删除它们 - prophet对丢失数据是不影响的。如果将历史数据的值设置`NA`为但在`future里`保留对应的日期，则Prophet将能提供其值的预测。

df.loc[(df['ds'] > '2010-01-01') & (df['ds'] < '2011-01-01'), 'y'] = None
    model = Prophet().fit(df)
    fig = model.plot(model.predict(future))

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzMzODczNDMx_size_16_color_FFFFFF_t_70 1][]

在上面的例子中，异常值扰乱了不确定性估计，但没有影响预测`值yhat`。情况并非总是如此，例如在以下示例中添加了异常值：

df = pd.read_csv('./data/example_wp_log_R_outliers2.csv')
    m = Prophet()
    m.fit(df)
    future = m.make_future_dataframe(periods=1096)
    forecast = m.predict(future)
    fig = m.plot(forecast)

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzMzODczNDMx_size_16_color_FFFFFF_t_70 2][]

此例中，在2015年6月造了一组极端异常值的数据来扰乱季节性估计，因此它们将影响到预测的值。和上个例子一样，正确的方法是删除它们：

df.loc[(df['ds'] > '2015-06-01') & (df['ds'] < '2015-06-30'), 'y'] = None
    m = Prophet().fit(df)
    fig = m.plot(m.predict(future))

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzMzODczNDMx_size_16_color_FFFFFF_t_70 3][]

参考资料：

[https://facebook.github.io/prophet/docs/outliers.html][https_facebook.github.io_prophet_docs_outliers.html]

[https_github.com_lilihongjava_prophet_demo_tree_master_outliers]: https://github.com/lilihongjava/prophet_demo/tree/master/outliers
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzMzODczNDMx_size_16_color_FFFFFF_t_70]: /images/20211104/adffdbcbf96945fb80b4caa77a269069.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzMzODczNDMx_size_16_color_FFFFFF_t_70 1]: /images/20211104/aa139602e8ee4427917b662bd4dd2861.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzMzODczNDMx_size_16_color_FFFFFF_t_70 2]: /images/20211104/f8f23717e1034b3fb93a598f9047a0e5.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzMzODczNDMx_size_16_color_FFFFFF_t_70 3]: /images/20211104/8d741ac9c4ce4fea9894f4013f3023c0.png
[https_facebook.github.io_prophet_docs_outliers.html]: https://facebook.github.io/prophet/docs/outliers.html