Zhe Sun is currently a senior data scientist in ING Wholesale banking Advanced Analytics team, where he has applied machine learning techniques to problems ranging from entity matching to large scale payment transaction network analysis. Together with the team, he aims to change the way the bank operates via data driven analytics and machine learning. He has 9 years of industry experience within data science and software engineering across a range of international companies, within the Banking and Telecommunications sectors.
June 5, 2018 05:00 PM PT
ING bank is a Dutch multinational, multi-product bank that offers banking services to 33 million retail and commercial customers in over 40 countries. At this scale, ING naturally faces a multitude of data consolidation tasks across its disparate sources. A common consolidation problem is fuzzy name matching: given a name (streaming) or a list of names (batch), find out the most similar name(s) from a different list.
Popular methods such as Levenshtein distance are not appropriate because of the time complexity and sheer volume of names involved. In this talk, we will introduce how we use a Spark custom ML pipeline and Structured Streaming to build fuzzy name matching products in batch and streaming. This can successfully match 8000 names per second against a 10 million name list, using a ten-node cluster. Firstly, we will give an introduction into the name matching problem.
Secondly, we will explain why Levenshtein distance approach is limited, and demonstrate a faster approach; token-based cosine similarity matching. Next, we will show how a ML pipeline helps to build an elegant solution. Here, we will deep dive into the detail of each stage, including customized preprocessing, tokenization, term-frequency, customized inverse document frequency, customized cosine similarity with distributed sparse matrix multiplication, and a customized supervision stage.
Finally, we will show how we deploy the ML pipeline within a batch data pipeline, and additionally as a fuzzy search engine in a streaming manner. Â The main conclusions will be: (1) a spark custom ML pipeline provides a powerful way to handle complicated data science problems (2) a uniform ML pipeline can serve both batch and streaming products easily from the same codebase.
Session hashtag: #MLSAIS17