Multimodal representation learning for retail product ontology

在过去的十年里，零售商经历了信息的大量激增，从他们的产品如何销售、他们的客户喜欢哪些产品，到他们的促销活动如何进行。数据可用性的这一巨大变化主要是由于物联网和数据仓库的进步，使数百万客户交易记录在大数据中心。然而，这些丰富的信息并不能轻易地转化为零售商和客户的利益，除非对这些信息进行严格的分析以指导商业决策。毫不奇怪，大数据分析现在正成为零售领域的业务需求，越来越多的零售商应用数据挖掘技术来改变他们的流程，从产品推荐到销售预测。尽管如此，零售商可以获得的大量原始数据是不完整的，也是通用的，无法证明有意义的分析是合理的。例如，许多产品仅以类型、价格、颜色等为特征，而告诉产品是否对特定年龄更有吸引力或是否更奢侈的信息通常是不存在的。这完全取决于人类专家的手动努力来识别这些“隐藏”属性，但这会给零售商或提供大数据分析服务的任何第三方带来巨大的成本和延迟。这项工作的目标是开发基于机器学习技术的新方法，以最小的人为干预自动发现这些隐藏的属性。这些方法将通过挖掘文本和推理与每个产品相关的图像来学习非显而易见的信息，结合自然语言处理和计算机视觉技术。通过这个过程，人们可以产生丰富的产品本体，从而实现高效的市场驱动分析。

In the past decade, retailers have experienced a massive surge in information ranging from how their productssell and which products their customers prefer, to how their promotional campaigns perform. This dramaticchange in data availability was mainly due to advances in the Internet-of-Things and data warehousing,allowing millions of customer transactions to be logged in big data centers. This abundance of information,however, cannot easily translate into benefits for retailer and customer, unless it is rigorously analyzed to guidebusiness decisions. It comes as no surprise that Big Data analytics are now becoming a business imperative inthe realm of retail, with more and more retailers applying data mining technologies to transform theirprocesses, from product recommendations to sales forecasting. Still, a large amount of the raw data available toretailers is incomplete and generic to justify meaningful analysis. For example many products are onlycharacterized by type, price, color etc., whereas information that tells if the product is more appealing tospecific ages, or that is more of a luxury product or not, is usually absent. It is entirely up to the manual effortof human experts to identify these "hidden" attributes, but this incurs significant costs and delays to the retaileror any third party that offers services on big data analytics. The goal of this work is to develop novelmethodologies based on machine learning technologies to automatically discover these hidden attributes withminimal human intervention. These methods will learn non-obvious information by mining text and reasoningaround images associated with each product, combining techniques from Natural Language Processing andComputer Vision. Through this process one can produce enriched product ontologies that allow for efficientmarket-driven analytics.