Customers nowadays expect more from their online store than just standard suggestions on the latest bestsellers or the cheapest items. Whoever wants to increase sales in e-commerce must meet the expectations and tastes of the customers - ideally fulfilling their wishes before they could even express them. However, this only works if the recommendations include comprehensive information such as the previous buying behavior and previous search queries, and thus become relevant for the customer. Pseudo-personalized emails, incoherent advertising, and randomly chosen "You might like it" links are more harmful than doing any good. If the personalized recommendation is successful, online retailers can improve customer satisfaction, achieve a higher conversion rate, and increase sales.

Two ways to personalized recommendation

The determination of personalized product recommendations usually takes place via two methods: A content-based system analyzes information about the potential buyer, such as their buying and browsing behavior. If there are matches, the corresponding product recommendation follows. If a customer's download list is full of books, e.g. by Isaac Asimov, Philip K. Dick or Arthur C. Clark, the online store recommends other Science-Fiction writers. Added value is provided by collaborative filtering, which incorporates the behavioral patterns of other buyers with similar interests when recommending a product.

If the sci-fi enthusiast exchanges, for instance, views on the latest novels with others in the forum, depending on the rating, his recommendations end up as well on the recommendation list of the panelists. So, one feature (genre sci-fi) is in the foreground, the other is relationships (client A - client B).

Regardless of the method, online providers need to be able to collect, contextualize and evaluate a wide variety of information. A complex task, considering the enormous amount of heterogeneous data and different data sources. In particular, the performance of the data analysis becomes the key factor, not to mention that every millisecond counts on the Internet. Hence, the query of the data must run in real time, if providers want to prevent frazzled customers migrate to the competitor. For conventional relational databases, which compute relationships through complex joins of primary and foreign key tables, such speed is difficult to achieve. Especially with large networked data, the query just takes too long. The recommendations might be based on outdated or incomplete information, so that it can happen that a customer is proposing a product that he has long bought, returned, complained or even reviewed negatively.

Relationships in focus with graph databases

Graph technology has established an alternative to traditional systems that map networked data to customers, products and services. This way, not only individual records but also the relationships between these entities can be retrieved. All the relevant information about a customer, including their social channels, search queries, and data from CRM systems and the online store are brought into an overall context that allows to make real-time, relevant and personalized recommendations.

The principle is simple: The graph model represents a set of objects along with the connections between those objects. The objects are called nodes, the connections between the nodes are called edges. Both nodes and edges can be assigned any number of properties and the links can be queried again, e.g. the price, rating, and genre of an article, or how long a product has been on a "watch list". A look at the example of the sci-fi fan makes the model clearer: customer A and book XZ are represented as nodes that are linked together by edges (e.g. "buys", "rated"). The line "rated" can be given the attribute "positive", "negative" or "neutral". If you want to use this information for referrals, you follow customer A's connections to other customers, identifying similar likes and possible referrals.

Graph databases were designed just for this type of query. Depending on the application, they can work up to a thousand times faster than conventional systems. Browser behaviors, search queries, click histories and wish-lists, as well as the additional connection with data from social channels are linked to a user profile that gives a 360 ° view of the customer. With every click and every purchase, this profile continues to grow. New customers, pre-marked articles, as well as comments and assessments are added and immediately considered in the next recommendation. So, the data record remains current.

In e-commerce, graph-based recommendation engines are used in the classic online shop as well as in comparison portals or for hotel and flight bookings.

Continuous data evaluation improves the algorithms, which further optimizes the product suggestions. The fact is that linking data is becoming more and more essential to keeping customers amid hundreds of thousands of products, offerings, and content over the long-term, and to encourage them to ultimately buy. Only with a timely and holistic analysis of customer and product data, online providers can improve the customer experience, strengthen loyalty to the company and thus increase their own online sales.

By Daniela La Marca