以基於時間比重之回歸預測電視收視率

Abstract

此論文主要貢獻為提出一個簡單且實驗結果準確的電視收視率預測方法，名為 Time Weighting Regression (TWR)。基於「越新的資料對預測接下來的收視率越重要」的假設，TWR 主要做的事情為：根據資料的時間賦予權重，再以帶有權重的資料建立回歸模型，最後用建立的模型預測接下來的收視率。我們以真實世界的電視收視率資料進行實驗，結果顯示它的預測比知名的時間序列模型（例如 Exponential Smoothing 和 ARIMA）和回歸模型（類神經網路）還準。

In this thesis, the primary contribution is proposing a simple and experimentally accurate solution, named Time Weighting Regression (TWR), to the problem of TV ratings prediction. Based on the assumption that newer data are more important for predicting upcoming ratings, what TWR does is: weighing data based on time, and then using weighted data to build regression model for predicting upcoming ratings. In the experiments on a real-world TV ratings data set, it outperforms well-known time series models (e.g., Exponential Smoothing and ARIMA) and regression model (neural network).