Mining the Deep Web using Sampling and Deep Learning Techniques

点击翻译按钮获取中文摘要

The deep web (or the hidden web) is the web that is hidden behind searchable interfaces. Unlike the surface web where pages can be browsed and hence downloaded in large scale, the access to the deep web is restricted. One common restriction is by queries via programmable Web APIs and web services. Many data sources, such as online social networks (OSNs), are examples of the deep web. They have a vast amount of data, but their access interface is restrictive and limited. Typically, they impose a quota for the queries we can send and data items we can retrieve per IP address. Discovering properties and patterns of the data hidden in the deep web is a challenging problem. Deep learning has been proven an effective approach to data mining tasks. It has been particularly successful in learning embeddings, i.e., short and dense continuous vector representations, for a variety of entities such as words, networks, and documents. Embeddings are essential for downstream data mining and machine learning tasks, such as classification, clustering, and recommendation. Embedding algorithms are data--hungry. Their success hinges on the availability of copious and pertinent training data. With the deep web, the training data are scarce, and may not be representative. We need to develop sampling techniques that can obtain pertinent data, and improve deep learning algorithms that can utilize the limited data. Humans learn not by reading all the text indexed by Google or GoogleScholar. Instead, we learn by sending pertinent queries, reading the returns, and sending new queries. Similarly, deep learning algorithms cannot and should not have all the text from Google or the entire social network from Facebook. Instead, there should be sampling-based deep learning algorithms that will learn from the deep web in an iterative process. The proposed research will approach the problem from two directions 1) Bottom-up from the deep web: we will study the sampling techniques that can be supported from real deep web sites such as Twitter; 2) Top-down from the deep learning: we will select several deep learning algorithms to study whether they can be approximated using samples from the deep web, and what kind of samples can improve the performance. We will start with neural network based representation learning, e.g., the state-of-the-art SN (Skipgram Negative Sampling) for text embedding and DeepWalk for graph embedding. After word embedding and node embedding, we will expand to document, linked document embedding, and author embeddings. The study will be conducted in two stages. In the first stage, we will evaluate our methods on our local academic search engine so that parameters can be controlled and ground truths are available. In the second stage, we will move on to real hidden data sources.