50 percent predictable: online shops can know well in advance what a customer will buy
The ability to predict customers' purchasing behavior gives online shops a lot of sway. Thanks to various analysis tools available today, companies can predict purchasing decisions in advance, making it possible to influence users' behavior, targeted recommendations, and, according to research from the 2015 SysRec Conference, predict with a 50% probability if the actual user session will end with a purchase or not.
Developed algorithms and the right technology can be an instrumental key toward e-commerce success. Through the analysis and evaluation of consumer behavior, clear predictions can be made about purchase times, especially involving recurrent purchases such as groceries, drugstore items and other consumer goods. This transparency is gaining in other categories as well, such as fashion, electronics, and travel. Big data analytics has made it clearer what customers like, as well as what they buy. It has become increasingly easier to make personalized offers, in order to give the consumer a gentle nudge and bring about a higher conversion rate.
This posts looks into how e-commerce stores can create a personalization system that delivers the exact information that will please customers and boost sales.
Every second purchasing decision is predictable
The ACM Recsys Challenge 2015 has shown that approximately 50 percent of all purchase decisions can be predicted in advance. Around 500 teams from 49 countries took part in the challenge of successfully predicting whether a customer would buy and what they would buy, by calculating an online merchant's actions based on past user behavior.
The ACM Recsys Challenge 2015 has shown that approximately 50 percent of all purchase decisions can be predicted in advance.
The making of personalization system is a multistage process, as detailed below.
Tracking codes and databases
First, enough data must be collected: all of a visitor's actions will be recorded as comprehensively as possible, with the help of a tracking code.
The result is a click stream, which contains information on every interaction with products and shop elements: what is the visitor's location, and what sort of device is the visitor using? When was the detailed product page viewed, and when was the product placed in the shopping cart or removed from it? What product was finally purchased, and at what price? Did the user click on the product's video or read the accompanying customer reviews?
This information can be captured with full respect to privacy. Personal data does not have to be recorded - a unique session ID is enough.
Larger online shops can quickly accrue millions of data records per month. NoSQL databases have been increasingly used for storage. These are helpful for collecting huge amounts of data quickly and cost-efficiently. With a long storage period, they enable powerful comparative data analysis. This wealth of historic data is the fuel that drives machine learning, and it can offer a glimpse into the future.
Making the most from feature engineering
Classic monitoring tools are not enough for making precise forecasts about purchasing decisions. Far more crucial to the process is the interim step called feature engineering.
During this process, data analysts examine the recorded data to find the key features within that data that allow for successful prediction. In the first step, this could be several hundred features - the time of day, or day of the week, for example; the amount of time that lapses between two visitor actions and the information gleaned from them; how long the visitor looked at the detailed product page; and so on.
Combinations of different actions are also often useful. For example, when visiting a product page more than once during a session, or at the start or end of a session; whether a product was placed in the shopping cart and whether this product was then removed.
Ten to twelve features are enough
The challenge for online merchants lies in identifying the most relevant features. These must be filtered out from the mass of evaluated raw data - do they deliver the desired result, or don't they?
In principle, many thousands of feature combinations are possible. Intelligent methods are necessary to derive exact prognoses from the recorded user data. As a rule, only about a dozen features should be used for the prognoses. In the Recsys Challenge, for example, the user's purchase intention was used to compute a 50 percent success rate in purchasing predictions.
Technological challenges and solutions
The prognosis models are prepared only once a day and stored in a fast database. Each calculation of concrete product recommendations takes place in milliseconds. While the visitor is navigating the shop, his or her actions are being tracked and compared with the prognoses, and from this the latest personal recommendations are shown. This can influence the order of future search results or even the appearance of a personalized website. Filtering functions can help with fine tuning and delivering more exact results.
Software as a Service
The use of NoSQL or in-memory databases, in connection with high-performance servers, is necessary for effective personalization.
Here is where cloud-based Software as a Service (SaaS) comes in. Online merchants can get relevant results without having to deal with databases and servers, as data collection and evaluation becomes increasingly an automated process. Merchants are given an API with which they can retrieve personal recommendations and search queries through tracking. Any online shop can therefore profit from personalization.
If you're considering using a personalization engine on your shop, or are interested in alternatives to your current recommendation service, take a look at eZ's cloud service eZ Personalization. Based on the YOOCHOOSE recommendation engine, eZ Personalization helps merchants increase conversion rates, build customer loyalty and increase your revenue.