A focused crawler aims at discovering as many web pages relevant to a target topic as possible, while avoiding irrelevant ones. Reinforcement Learning (RL) has been utilized to optimize focused crawling. In this paper, we propose TRES, an RL-empowered framework for focused crawling. We model the crawling environment as a Markov Decision Process, which the RL agent aims at solving by determining a good crawling strategy. Starting from a few human provided keywords and a small text corpus, that are expected to be relevant to the target topic, TRES follows a keyword set expansion procedure, which guides crawling, and trains a classifier that constitutes the reward function. To avoid a computationally infeasible brute force method for selecting a best action, we propose Tree-Frontier, a decision-tree-based algorithm that adaptively discretizes the large state and action spaces and finds only a few representative actions. Tree-Frontier allows the agent to be likely to select near-optimal actions by being greedy over selecting the best representative action. Experimentally, we show that TRES significantly outperforms state-of-the-art methods in terms of harvest rate (ratio of relevant pages crawled), while Tree-Frontier reduces by orders of magnitude the number of actions needed to be evaluated at each timestep.
翻译:集中的爬行器旨在尽可能多地发现与目标主题相关的网页,同时避免不相干的内容。强化学习(RL)已被用于优化重点爬行。在本文中,我们提议TRES,这是一个有重点爬行的RL动力框架。我们把爬行环境模型成一个Markov 决策程序,RL代理商旨在通过确定一个良好的爬行战略来解决这个问题。从几个与目标主题相关的人类提供的关键字和一个小文本体开始,TRE遵循一个关键字集扩展程序,该程序引导爬行,并训练一个构成奖赏功能的分类师。为了避免一种计算上不可行的粗力方法选择最佳行动,我们提议了树-Frontier,一种基于决定的算法,它适应性地将大型状态和行动空间分散,只找到少数具有代表性的行动。树-Frontier允许该代理商通过贪婪地选择最佳的代表性行动来选择接近最优化的行动。实验,我们显示TRES明显地超越了选择最先进的标准状态,同时按不同程度的收成品级排序,同时评估每一程度的顺序。