Using Machine Learning Methods to Classify and Retrieve Web APIs

• Main Authors: Pan, Tian-Yun, 潘天允
• Other Authors: 馬尚彬
• Format: Others
• Language: zh-TW
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/65m3ch

碩士 === 國立臺灣海洋大學 === 資訊工程學系 === 106 === Nowadays, the retrieval of Web APIs to build applications is a must for developers. However, existing API systems are not able to help developers to precisely locate Web APIs. We identified two weaknesses of current API systems: 1) API category is an important way to find relevant APIs. Current systems usually classify APIs into only one category. The single-category design is not reasonable and hinders the retrieval of relevant Web APIs, since most APIs belong to multiple categories; 2) Current systems expect users to issue keywords when searching APIs, not considering semantics. It makes that the search results are very likely few or unrelated. Therefore, in this thesis, we propose a new system, referred to as API Locator, to address the above issues. In API Locator, we apply the techniques of Machine Learning and most commonly used NLP (Nature Language Processing) mechanisms to devise a set of classification and recommendation algorithms. These algorithms are divided into three parts: 1) classification identification, which applies the traditional machine learning (SGD Classifier) and deep learning (LSTM) to identify the classification tags; 2) classification exclusion, which uses Word2Vec to exclude unrelated tags semantically; 3) API recommendation, which retrieves and sorts all the possible APIs with TF-IDF and Cosine Similarity. Notably, unlike the traditional search mode, the API Locator provides users with a way of searching by descriptions and supplies more possible classification information to the user query. Besides, the experimental results show high precision to multi-classification and API retrieval and demonstrate the feasibility and efficacy of API Locator.