Matching product titles using web-based enrichment

Abstract

Matching product titles from different data feeds that refer to the same underlying product entity is a key problem in online shopping. This matching problem is challenging because titles across the feeds have diverse representations with some missing important keywords like brand and others containing extraneous keywords related to product specifications. In this paper, we propose a novel unsupervised matching algorithm that leverages web search engines to (1) enrich product titles by adding important missing tokens that occur frequently in search results, and (2) compute importance scores for tokens based on their ability to retrieve other (enriched title) tokens in search results. Our matching scheme calculates the Cosine similarity between enriched title pairs with tokens weighted by their importance scores. We propose an optimization that exploits the templatized structure of product titles to reduce the number of search queries. In experiments with real-life shopping datasets, we found that our matching algorithm has superior F1 scores compared to IDF-based cosine similarity.

Matching titles with cross title web-search enrichment and community detection

“…However, these basic similarity functions cannot consistently perform well across the remaining three categories. Variations of the basic similarity approach, such as "Term Frequency × Inverse Document Frequency" (TF-IDF) [22] that has a global scope and stable ranking mechanism, can handle Category 3 fairly well, but fail in Category 4 for the reasons presented in [16].…”

“…There have been some recent works such as [8] and [16] (EN+IMP) that try to overcome the above limitations of these basic approaches. Although these fare better than most standard approaches -see Category 4, their efficacy is limited.…”

“…Although these fare better than most standard approaches -see Category 4, their efficacy is limited. This is largely due to their dependency on reducing the comparison to exact matches [8] or using a similarity measure that makes use of weights that are computed in a way which is agnostic to the compared title [16]. Furthermore, both these methods, as well as TF-IDF, fail in understanding the semantic relationships -see TF-IDF and EN+IMP for Category 2 in Table 1.…”

“…The essence of this problem has now blended itself in a variety of modern settings in the Internet-age, including aligning multilingual texts for machine translation, IP aliasing in networks [25], reconciling photographs with correct tags [21], or matching product titles across different online merchants [8,16,18]. The diversity in the target domains has led to a varied set of solutions tailored for individual domains.…”

“…For example the product "d200 10.2 megapixel digital slr camera body with lens kit -18mm -135mm (2.5"lcd -7.5x optical zoom -3872×2592 image)" from PriceGrabber.com is represented simply as "nikon d200" at CNET.com [16]. This problem is very common, in part due to the general lack of standardisation in how a product is published, curated and managed [17].…”