運用長短期記憶模型及對比學習於產品生命週期預測

Abstract

產品生命週期（product life cycle）通常被作為分析產品發展軌跡的基礎，它幫助了生產、行銷和定價等關鍵決策的制定。精確的產品生命週期預測為管理者提供 了預測未來產品銷售量的優勢。然而，傳統的統計方法存在局限性，主要在於它們無法適應多樣化的應用環境。相比之下，深度學習（deep learning）方法為提高產品生命週期預測的準確性和適應性提供了一個有潛力的途徑。在本論文中，我們通過引入表徵學習（representational learning）中的對比學習（contrastive learning）並與長短期記憶（LSTM）模型相結合，構建了先進的產品生命週期預測模型，以填補這一領域的研究空白。我們的研究對產品生命週期預測領域有著多方面的貢獻。首先，我們的實驗證明了對比學習模型在增強時間序列預測方面的有效性，我們的對比學習模型能夠有效地從原始數據中提取重要資訊。其次，我們提出了多種將對比學習與長短期記憶編碼-解碼模型相結合的方法。我們提出的模型優於傳統方法，包括基本的長短期記憶模型，從而確鑿地展示了其提高產品生命週期預測的能力。

The Product Life Cycle (PLC) serves as a foundation for understanding a product's journey in the market, guiding crucial decisions in production, marketing, and pricing. Precise PLC forecasting offers managers the advantage of anticipating future product sales volume. However, conventional statistical methods exhibit limitations, primarily due to their reliance on extensive training data and their inability to adapt to diverse industrial contexts. In contrast, deep learning approaches present a promising avenue for elevating the accuracy and adaptability of PLC forecasts. In this thesis, we bridge this gap by introducing contrastive learning alongside a Long Short-Term Memory (LSTM) model to craft an advanced PLC forecasting model. Our research contributes to the field of PLC forecasting in two ways. Firstly, we underscore the effectiveness of contrastive learning embeddings in enhancing time series predictions, with our contrastive model effectively extracting vital information from raw data. Secondly, we propose various approaches integrating contrastive learning with the LSTM encoder-decoder model. Our proposed models outperform traditional methods, including basic LSTM models, conclusively demonstrating their potential to enhance PLC forecasts.