API Code Recommendation using Statistical Learning from Fine-Grained Changes

來自專欄 API推薦及代碼補全

本次分享的文章是2016年軟工頂級會議FSE的一篇論文。In summary, 技術思路還較為常規，但實驗部分非常突出。

作者：

• Tien N. Nguyen, Associate Professor, Department of Computer Science, Software Engineering Program, low State University
  Research interest: Software engineering: 1) Program Analysis and Understanding , 2) software Evolution and Maintenance, 3) Mining Software repositories, 4) Version and Configuration Management, 5) Web and Configurable Code Analysis

• Danny Dig. associate professor, EECS Department, Oregon State University
  Research interests: automated refactoring, interactive program analysis and transformation, design & architectural patterns, and broadly interested in concurrency and parallelism. mobile computing, software development processes, program analysis, softe=ware testing, and software evolution.

背景介紹：

the state-of-the-art research in code completion

• Using API usage patterns：

  1) mine via the deterministic algorithms such as frequent itemset mining, pair ass-ociations, frequent subsequence or subgraph mining.

  2) to find the matching pattern, then recommend the rest of the API elements in the pattern

• Using statistical learning via language models:
  1) rely on the regularity of source code and create a model that statistically learns code patterns from a large corpus.
  2) can predict what token is likely to follow a sequence of given code elements.

  3) Limitation: difficult to determine which tokens belong to a project-specific code idiom, which produces noise.

本文中：

They present a novel approach to code completion that leverages the regularity and repetitiveness of software changes. We also consider the surrounding code context at the recommendation point.

技術簡介：

Key ideas：Repetitiveness of software changes

• Instead of relying on source code repetitiveness, APIREC is based on code change repetitiveness. Grouping fine-grained changes by higher-level intent allows us to cut through the noise of unrelated tokens.
• The changes that belong to higher-level intents will co-occur more frequently than non-related changes. And we use associations among fine-grained changes so that changes neednt have strict order as required in n-gram model.

實驗介紹：

未完待續 T_T

推薦閱讀：