- Comprehensive E-Commerce platform.
- Cross-manufacturer product data overflow.
- Channel-wide product feed standard absent.
- Datasets vary in formats, atributes, structures.
- Various levels of data quality and accuracy.
HOW ONEDOT HELPED
- Automatically matched extensive product categories.
- AI-algorithm matched vendor products.
- Key attributes were normalised.
- Product duplicates identified.
- Unnecessary database content cleared.
STORE is an online marketplace for hardware and software as well as consumer electronics, household goods, sports and leisure articles. With offices in Zürich, Bern, and Berlin, Germany, STORE is currently comprising about ten million products and established itself as a unique online marketplace.
The online marketplace STORE exceeded its planned growth trajectory for the year 2018, mainly because of the company’s customer-centric culture. The rise of the popularity and their dedicated team of accessible sale experts work directly with customers.
STORE’s database encompasses thousands if brands and 5+ million stock keeping units, making STORE the #1 online marketplace for consumer electronics, household goods, sports and leisure articles. High-quality product data and attributes are crucial for better sales today.
The challenge: STORE strives for scalable data processing, especially in the area of schema mapping and attribute normalization in order to reduce manual effort and achievie time savings. Additionally, it is crucial to improve data quality and increase conversion rate. This needs a scalable data preparation to reach faster more SKUs online.
The main line is to shorten the time-to-market, increases conversion rate and reduce operational costs. Incoming vendor datasets and the existing STORE data-base have different formats, attributes, structures and also various levels of data quality and accuracy making it difficult to integrate.
“Onedot supports STORE to grow customer satisfaction by automating vendor onboarding, making products more searchable as well as accessible to improve information quality within the expanding database.”
Dr. Stephan Weber, CRO at STORE Inc.
Onedot’s solution includes detailed data preparation, algorithmic data wrangling, and dataset management. Key value pairs were extracted from individual attributes to reach one attribute per column. Scheme Mapping and Data Integration of the source attributes per customer request and per catergory. And then normalization of structured product information could be reached. In detail:
1. Extraction: splitting of key value pairs in into Attribute and Attribute Value, structuring of extracted attributes with one attribute per column. Scaled data was available after one day.
2. Mapping and integration starts with no corresponding attribute found in input data, followed by one on one mapping of input attribute to target schema, then merged as set. After that one Input attribute has been split and mapped to several attributes of the target schema. This reached 71% destination attributes filled with one plus input attributes.
3. Normalization: best practice approach and customer specifications normalize and convert attribute values. Example: Replace “ß” with “ss” (CH notation)
“Onedot used artificial intelligence (AI)-powered data preparation within a short period of time to solve the issue. This fast service automatically integrates, transforms and enriches data from various sources of structured and semi-structured data, even exceeding human precision.”
Dr. Stephan Weber, CRO at STORE Inc.
The STORE will significantly benefit from the increase in data quality and categorisation accuracy. Locating online products faster and more precisely will enable STORE to take the next steps in customer centricity while delivering the best possible search results. Automatically adding and categorising thousands of products speeded up the on-boarding process and the go-live of the new big vendor, drastically.
With a precision and recall of over 95%, the incoming vendor products were successfully categorised into over 800 categories. Plus, only a few iterations and minimal training data were required to train the categorisation classifier. Golden records were identified using fuzzy matching across predefined product attributes.