The product domain represents a challenging scenario for developing and evaluating big data integration solutions: the number of sources providing product speci cations is very large, and ever increasing over time. The volume of available data is impressive, and these data keep changing very frequently. In this paper, we present ongoing e orts, challenges and our research agenda to address big data integration for product speci cations.
This paper describes our recent experiences dealing with the issue of extracting and integrating data from product speci cation pages on the Web, the challenges that we encountered and opportunities we leveraged. We provide an overview on our ongoing research activities on this problem, we discuss the lessons that we have learned from our experience, and illustrate open problems and future directions.
Integrating data from the product domain represents a challenging issue, presenting most of the di culties associated with big data solutions at the web scale. Thousands of websites contain an impressive number of product pages. Each page provides information about a single product: usually general product speci cations (typically technical and physical features), and some source-speci c data, such as, price and customers' reviews.
Integrating product data might enable many valuable applications, such as
data driven market analysis, question answering, and price comparison. However,
data integration at the web-scale raises intriguing challenges due to the volume,
variety, velocity, and veracity of data [
To address the above challenges, we are working on an end-to-end data integration solution for product speci cation pages. For simplicity of presentation, we describe our approach as a linear pipeline, as depicted in Figure 1, where tasks are performed in sequence and independent of one another. However, there might be feedback loops between the tasks, as intermediate results can indeed in uence the performance and the behavior of the preceding tasks and of the end to end solution.
In our vision, the information need is expressed by an input set of sample
pages. We observe that products are typically organized in categories, and hence
we expect that the sample pages refer to products from the categories of interest.
Our approach is inspired by the Open Information Extraction [
Every task produces output to feed the successive task in the pipeline, but
intermediate results could nd other compelling application scenarios as well.
To this end, we advocate that the pipeline must include an empirical evaluation
benchmark for every task.
Source
Discovery
Evaluation
Evaluation
Evaluation
Evaluation
Evaluation
Source discovery aims at e ciently nding and crawling product websites in
order to gather pages that refer to the products of interest. One might believe
that discovering product websites is a minor task, as a relatively small number
of head sources can o er enough data for most of the products. For example,
amazon.com already provides data about an impressive number of products.
However, valuable information is actually published by an enormous number
of tail sources [
Data discovery and extraction has the objective of processing the pages harvested in the previous task in order to locate and extract product attribute names and their values. As we mentioned above, we do not rely on a prede ned schema, but rather extract attributes bottom-up, with the goal of discovering not just head attributes, but also tail attributes that cannot always be described in advance.
Data linkage seeks to cluster pages from di erent sources that refer to the same
products. It is worth observing that in the traditional data integration pipeline,
schema alignment is performed before record linkage [
Schema alignment addresses the challenge of semantic ambiguity and aims to reconcile the attributes o ered by di erent sources, that is, to understand which attributes have the same meaning and which ones do not, as well as identify value transformations to normalize di erent representations of the same attribute values. Since we do not rely on a global schema given in advance, correspondences among attributes are established bottom-up leveraging the results of the previous data extraction and data linkage phases. Data fusion tackles the issue of reconciling con icting values that may occur for attributes from di erent sources. Data fusion aims at evaluating the trustworthiness of data, deciding the true value for each data item, and the accuracy of the sources. To address these challenges, data fusion techniques rely on data redundancy, which further motivates the need to process many sources. 2.1
Our approach: redundancy as a friend Our approach to develop the above pipeline aims at taking advantage of the opportunity that products are named entities, and hence a product speci cation page usually publishes the product identi er. Web sources that deliver product speci cation pages publish product identi ers mainly for economic reasons: websites need to expose the product identi ers to let them be indexed by shopping agents and available to customers who search products for comparing prices or consulting speci cations. Large e-commerce marketplaces strongly encourage sellers and retailers to publish product identi ers, as they improve e ciency both for the internal management of data and for the exchange of data with search engines like Google and Bing.
The presence of identi ers allows us to drive the pipeline from source discovery to data integration by leveraging the opportunity of redundancy of information at the global level, and the homogeneity of information at the local level.
{ At the global level, we observe that head (popular) products are present in several head (large) sources as well as in many tail (small) sources. Therefore, we expect that identi ers of head products are spread across many sources. Further, many head products in a category will often co-occur in multiple sources. { At the local level, we observe that the structure and the semantics of information, within each source, tend to be regular. Hence, we expect the product speci cation and product identi ers presented in a given page are published according to the same structure for every page in the same source.
Figure 2 illustrates the key intuitions underlying our approach from source discovery to data integration to meet the goal of e ectively and e ciently dealing with head and tail sources, hence including all pages of head and tail entities. Starting from known head entities in head sources, we take advantage of homogeneity of information at the local level to extract product speci cations and identi ers for tail entities in head sources (even head sources o er many tail entities). Then, we exploit the presence of head entities across sources: searching head identi ers, we discover tail sources (even tail sources o er a few head entities). Again, we exploit homogeneity of information at the local level to extract identi ers and speci cations for tail entities in tail sources.
eBay, Amazon, Google Shop explicitly require sellers to publish the id for many product categories. For example, see eBay's rules: http://for-business.ebay.com/product-identi ers-what-they-are-and-why-they-areimportant.
Based on our Redundancy as a friend approach, we developed a focused
crawler, Dexter [
In an experimental evaluation performed between Sept 2014 and Feb 2015, we
have trained the focused Dexter crawler [
Building a Benchmark Product Dataset { We compared the contents of our
dataset with pages in Common Crawl, an open repository of web crawl data.
About 68% of the sources discovered by our approach were not present in
Common Crawl. Only 20% of our sources contained fewer pages than the same sources
in Common Crawl, and a very small fraction of the pages in these sources were
product pages: on a sample set of 12 websites where Common Crawl presented
Note that the on-line version (https://github.com/disheng/DEXTER) is an
extension of the dataset presented in [
These results suggest the critical need for the community to build a suitable benchmark product dataset to conduct big data research. Our dataset can be considered a rst step in this direction. Another important step would be that of maintaining the dataset over time, as discussed next.
Maintaining the Benchmark Product Dataset: Addressing the Velocity Challenge { In March 2018, we have checked all the URLs of the pages of the dataset. We have observed that just 30% of the original pages and 37% of the original sources are still valid (we consider a source valid if it contains at least one working URL). We also performed an extraction of the product speci cations. We obtained complete speci cation from just 20% of the pages.
These numbers clearly indicate that the velocity dimension a ects all the tasks of the pipeline. Developing solutions to collect snapshots over regular time intervals and perform data integration over time can open intriguing research directions. While some activities, such as checking the appearance/disappearance of sources can be done on monthly basis, others, such as crawling web sites to check appearance/disappearance of pages and changes in the pages should be performed more frequently. To this end, the development of e cient incremental solutions for source discovery and web crawling represent interesting research directions.
As our experiments emphasize, data extraction rules are brittle over time. The development of wrappers resilient to changes in the pages has always been a primary goal in data extraction research. A dataset with multiple snapshots over a long interval of time, as the one that we have advocated above, could serve as a benchmark for data extraction solutions.
Harnessing Velocity { We observe that while velocity represents a challenge, it could also become an opportunity. We observe that analyzing changes in the pages could help improve our data extraction and data linkage techniques.
For example, one of the main challenges of the product identi er extraction step is to eliminate identi ers that refer to suggested and related products, instead of the main product of the page. Examining the same product page over time may help us to more easily separate out the former part, since those may change faster over time than the description of the product in the page. Schema Alignment { We are currently working on the development of techniques to perform schema alignment for our product domain. The main di culties that we have to face are due to the heterogeneity at the schema and at the instance level, due to the large number of independent sources.
To give a concrete example of the heterogeneity at the schema level, consider the dataset collected using the Dexter crawler, described earlier in this section. The speci cations extracted from these sources contain more than 86k distinct attribute names (after normalization by lowercasing and removal of non alphanumeric characters). Most of the attribute names (about 85k) are present in less than 3% of the sources, while only 80 attribute names occur in 10% of the sources, with the most popular attribute name occurring in just 38% sources.
The solution that we are investigating exploits data linkage to cluster attributes that share the same values. Since heterogeneity occurs also at the instance level, we aim at progressively nding correspondences between the cluster resolving increasingly complex matches between values with di erent representations.
Addressing the Veracity Challenge and Data Fusion { Analyzing the sources of our dataset we noticed that some clusters contain product pages from di erent categories. In some cases, the errors are in the sources: some websites adopt unexpected (or even completely wrong) criteria as, for example, classifying monitors under a laptop category or vice-versa; other websites aggregate products and related accessories (which represent another category of products). In other cases, the errors are due to wrong classi cations by the system.
To overcome these issues we want to investigate di erent solutions. First, we believe that introducing feedbacks in our pipeline could signi cantly improve data quality, especially for precision. For example, alternating source discovery and data linkage we could select better identi ers to feed to Search for source discovery, thus increasing the precision, with respect to the target category, of the sources. Similarly, results from schema alignment could help improve the precision of linkage, as pages whose attributes do not align are unlikely to represent the same product. Another promising direction to improve precision without penalizing recall is to study solutions to exploit humans in the loop. In particular, we aim at developing and evaluating techniques based on active learning and crowdsourcing to continuously train the classi ers with e ective and updated training sets.
We have observed many inconsistencies between values of a product attribute
across sources. Existing data fusion techniques [
Beyond Source Homogeneity { Our approach to data extraction is based on the assumption that pages are structurally homogeneous at the local source level. This is a valid assumption for a vast majority of websites, but there are exceptions. For example, some websites that publish used products have a weak template and leave the seller the freedom to publish the product speci cation without any structural imposition. For some application scenarios, one can simply drop these sources. If on the contrary they are important, data extraction should be performed by solutions that do not rely on the template. Beyond Product Speci cations { So far we have considered the extraction of the identi ers and of the speci cations. However important data that complete the product description are price and reviews. Challenging issues for the extraction of price are to distinguish the price of the principal product in the page from the prices of other products, such as suggested products, similar products, and the actual price from discounts or list price. Reviews represent important information in many applications. An interesting problem is how to combine structured data from the speci cation with the unstructured data of the reviews.