透過再預訓練BERT結合適配器與情感分析 預測線上平台遊戲產品存活年限;Predicting the Lifespan of Online Platform Games through BERT Re-Pretraining with Adapters and Sentiment Analysis

NCUIR > School of Management at National Central University > Graduate Institute of Industrial Management > Electronic Thesis & Dissertation > Item 987654321/97311

Please use this identifier to cite or link to this item: https://ir.lib.ncu.edu.tw/handle/987654321/97311

Title:	透過再預訓練BERT結合適配器與情感分析預測線上平台遊戲產品存活年限;Predicting the Lifespan of Online Platform Games through BERT Re-Pretraining with Adapters and Sentiment Analysis
Authors:	游家齊;Yu, Chia-Chi
Contributors:	工業管理研究所
Keywords:	遊戲壽命預測;BERT輕量模型;MLM 預訓練;評論分析;模型微調;Game Lifespan Prediction;BERT Model;MLM Pretraining;User Review Analysis;Model Finetuning
Date:	2025-07-21
Issue Date:	2025-10-17 11:07:20 (UTC+8)
Publisher:	國立中央大學
Abstract:	隨著數位遊戲產業蓬勃發展，大量遊戲產品不斷上架與下架，遊戲壽命逐漸成為玩家課金與廠商營運決策的重要參考依據。然而，現行平台多未揭露遊戲存續風險資訊，導致玩家可能投入於即將下架之遊戲，產生資源浪費與信任流失問題。因此，本研究旨在建立一套可預測遊戲存活年限之模型架構，協助使用者判斷遊戲潛在壽命長度，提升消費決策品質。本研究以 Steam 平台為資料來源，透過網路爬蟲蒐集已下架遊戲之上架與下架年份，計算遊戲存活年數作為標籤，並搭配遊戲評論文本進行預測建模。模型訓練流程包含：BERT 基礎模型之 Masked Language Modeling（MLM）預訓練，後續依序加入 Adapter 模組、情緒提示（Prefix）與完整參數微調（Finetune），建構逐層強化的預測模型。另設計驗證實驗，模擬模型僅能取得歷史資料時，對未來下架遊戲的預測能力。實驗結果顯示，模型表現隨模組設計漸進提升，其中結合 Adapter、Prefix 並進行完整微調之模型於準確率、平均絕對誤差（MAE）、±1 年與 ±2 年準確率等指標皆展現最佳表現。此外，跨時期驗證結果顯示，僅以 2023 年以前資料訓練之模型，其預測結果與資料量對照組相近，顯示本研究模型具備良好泛化能力，可應用於實務中對遊戲壽命進行預測與風險判斷。未來將進一步建置查詢介面，使使用者能輸入遊戲 ID，即可獲得預測年限區間與落點機率，實現實用化推廣目標。;With the rapid growth of the digital gaming industry, a vast number of games are being released and subsequently delisted. As a result, the lifespan of games has become a critical factor for players′ spending decisions and developers′ operational planning. However, current gaming platforms rarely disclose information regarding a game′s potential longevity, which may lead players to invest in games that are likely to be delisted soon, resulting in wasted resources and diminished trust. This study aims to construct a predictive model for estimating a game′s survival duration, thereby assisting users in making more informed decisions. Using Steam as the data source, this study collected the release and delisting years of games via web scraping and calculated each game′s survival duration as the classification label. Coupled with user review texts, a series of models were trained to predict the number of years a game would remain available. The modeling process includes: (1) Masked Language Modeling (MLM) pretraining using a BERT-based architecture, followed by (2) the integration of Adapter modules, (3) the inclusion of sentiment-aware prefixes, and (4) full-parameter finetuning to progressively enhance performance. A validation experiment was also designed to simulate the real-world scenario in which only historical data is available, to assess the model’s generalization to future game delistings. Experimental results show that model performance improved significantly with each additional enhancement. The final model—combining Adapters, Prefix, and Finetuning—achieved the best results in terms of accuracy, mean absolute error (MAE), and ±1 / ±2-year prediction accuracy. Furthermore, in the cross-period validation experiment, the model trained solely on data prior to 2023 achieved comparable performance to a baseline model trained on a random subset of the full dataset. This indicates strong generalization ability and practical potential for forecasting future game lifespans. Future work will involve the development of a user-facing prediction interface, allowing users to input a game ID and receive a predicted lifespan range along with the probability of the game falling within that range—thus facilitating the practical deployment of this model in real-world scenarios.
Appears in Collections:	[Graduate Institute of Industrial Management] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	22	View/Open

社群 sharing

Loading...