Product Recommendation Ststem Based Semi-Structured Information

Product recommendation is to select and recommend proper products to users from the massive product pools of online shopping web site. A good product recommendation system not only can greatly save user time of searching and selecting the needed products, but also can effectively increase the user stickiness for a trade. Therefore, product recommendation system has been widely used in electronic business, meanwhile attracted more and more researchers’ attention. However, it still exists many problems to be resolved.Based on existing research, this thesis focuses on two techniques to improving the product recommendation results based on semi-structured information, one is to extract structured product attribute information from semi-structured web texts, and the second is to optimize the similarity metric for product comparison based on the structured attribute information To this end, this paper carried out the following concrete research:(1)To efficiently extract the product attributes and attribute values, this thesis explores useful information for distinguishing a product from the semi-structured product description web pages, and then adopts a linear conditional random field model based on active learning to extract product attributes and attribute values. The experiments show that this method can obtain better extraction performance, meanwhile reduce the workload of manually annotating corpus, therefore, has a better portability.(2) To precisely compare products, this thesis proposes a similarity measure based on the extracted product attributes. This method first classifies products into different categories based on product attributes, and then minimum edit distance consolidated other attributes between same category products to measure product similarity. The experiments show that the effectiveness of the method.(3) To build a practical product recommendation system based on semi-structured information, this thesis designs and implements a prototype with JSP Servlet Spring Hibernate framework. The experiments show that the proposed method yields a better recommendation results than several strong baseline methods, and processes better interpretability and scalability.