Abstract:
重回帰分析により工数見積りモデルを構築する場合，何らかの基準に基づいてプロジェクトを層別（分類）したデータセットを作成し，分類ごとに見積りモデルを構築することが望ましい．本論文では，生産性と関連が強い（寄与率の大きな）プロジェクト特性（開発言語，アーキテクチャ等）を用いて層別を行うことを提案する．財団法人経済調査会が収集した153プロジェクトより抽出した，43プロジェクトの実績データを分析し，規模当たり要員数（平均要員数／FP）の寄与率が大きい（寄与率38%）ことが分かった．一般に見積りモデル利用時には（規模当たり）要員数は確定していないが，規模当たり要員数を概算値（大，中，小の3段階）で決めることは比較的容易である．規模当たり要員数を用いてプロジェクトを3つに層別した結果，見積り精度が大きく改善する（要員数概算値の相対誤差平均が50%の場合，工数見積り値の相対誤差中央値が56.6%から35.9%になる）ことが分かった．; When building effort estimation models by multiple regression analysis, it is preferable to stratify software projects according to some criteria, and to build models for each stratified projects. In this paper, we propose to stratify projects by project attributes (programming language, and architecture et al.) which have strong relationships to productivity when building effort estimation models. We analyzed 43 projects data selected from 153 projects collected by Economic Research Association, and concluded that team size per project size had strongest relationship to productivity (variance explained was 38%). Although in general size per project size is not fixed when estimating effort, it is not very difficult that size per project is fixed approximately. We stratified projects by team size per project size and built models for each stratified projects. As a result, the accuracy of the models showed better performance than a model built from whole projects; and, the median of relative error of estimated effort was improved from 56.6% to 35.9% when the average of relative error of estimated team size was 50%.
