Rec­om­men­da­tion systems in e-commerce

Rec­om­men­da­tion systems have gained im­por­tance in recent years, and appear very promising for e-commerce, es­pe­cial­ly in con­nec­tion with Big Data. By working with large amounts of data and using so­phis­ti­cat­ed al­go­rithms, you can increase the con­ver­sions of online stores with modern rec­om­men­da­tion systems. You are no doubt familiar with the classic 'cus­tomers who bought this item also bought…' or 'the following products may interest you…' no­ti­fi­ca­tions in online shopping. These tips, cor­re­spond­ing to the in­di­vid­ual pref­er­ences of the user, result from con­sid­er­able computing power and com­pli­cat­ed al­go­rithms. In e-commerce, rec­om­men­da­tion systems are already in­cor­po­rat­ed, but other sectors also benefit from con­tin­u­ous­ly improving al­go­rithms and tech­nolo­gies.

A rec­om­men­da­tion system (also known as 'rec­om­mender system', 'rec­om­men­da­tion engine', or 'rec­om­men­da­tion platform') is, according to the de­f­i­n­i­tion, an in­for­ma­tion filtering system that attempts to predict the 'rating' or 'pref­er­ence' that a user might give to an item or product. It analyzes past behavior and previous orders, and au­to­mat­i­cal­ly searches for similar and in­ter­est­ing products for the user. The field of ap­pli­ca­tion for rec­om­men­da­tion systems is diverse. You will recognize rec­om­men­da­tions from web stores, streaming services, and online pub­li­ca­tions – wherever large amounts of objects are, whether they are books, clothes, or movies – but only a small amount has any relevance to the user. Con­sid­er­ing the amount of data and possible search paths, rec­om­men­da­tions help by selecting a small amount in advance from the confusing selection and pre­sent­ing it to the user.

Ad­van­tages of rec­om­men­da­tion systems

These systems should make searching a lot easier for the user. Instead of clicking through lots of offers and pages to find the right product, this pre-selection aims to exclude ir­rel­e­vant and un­in­ter­est­ing offers so that the ones displayed are the ones most suited to the user.

Operators also hope for a positive effect, such as an increase in visitor numbers in the content sector or an increase in sales in e-commerce. In online stores, suitable sug­ges­tions should ideally lead to bigger shopping carts, which then sig­nif­i­cant­ly increase margins.

But cal­cu­lat­ing using al­go­rithms also comes with its dis­ad­van­tages. Often personal, human com­po­nents are missing when it comes to the suggested selection. Even so­phis­ti­cat­ed cal­cu­la­tions can read simple human behaviors wrong and are more likely to leave the user shaking their head rather than impress them. For example, Amazon might display a glass cutter, which many craftsmen need for their daily work, but the selection also shows bal­a­clavas and other ac­ces­sories more suited to burglars.

Rec­om­men­da­tion services are always based on sets of data. Depending on the nature of this data set, a dis­tinc­tion is made between different types of systems. These are typically content-based and col­lab­o­ra­tive systems. Fur­ther­more, there are also context-sensitive rec­om­men­da­tion services, as well as those that include the chrono­log­i­cal sequence or de­mo­graph­ic user data in the analysis.

Content-based rec­om­men­da­tion systems

Content-based rec­om­men­da­tion systems suggest objects or content similar to what the user has already searched for, viewed, bought, or rated highly. The system must be able to establish a sim­i­lar­i­ty between objects. This is done through content analysis. For music streaming services, for example, the system analyzes a piece of music (e.g. the structure of the music) to find pieces that have a similar bass track.

Col­lab­o­ra­tive rec­om­men­da­tion systems

When it comes to col­lab­o­ra­tive methods, the sug­ges­tions are based on users with similar rating behavior. If they showed a lot of interest in a par­tic­u­lar object in the past, the system continues to suggest it. In­for­ma­tion or knowledge about the object itself is not necessary at this point. Amazon, for example, uses this method ex­ten­sive­ly.

Different pre­dic­tion methods

Rec­om­men­da­tion services use different learning methods. In general, either the memory-based or the model-based method is used. The memory-based method uses all stored eval­u­a­tion data and cal­cu­lates the sim­i­lar­i­ty between users or objects. The result is the base, which enables user-object com­bi­na­tions to be predicted. Model-based rec­om­men­da­tion services, on the other hand, work with the prin­ci­ples of machine learning. Based on the data, the system can create a math­e­mat­i­cal model that can be used to predict the user’s interest in a par­tic­u­lar product.

Rec­om­men­da­tion systems can be found in many sectors and branches. The three most important are large streaming services such as Spotify or Netflix, classic e-commerce providers such as Amazon, and special rec­om­men­da­tion systems for content-based ad­ver­tis­ing.

Amazingly, Amazon released its deep-learning software DSSTNE as open source at the beginning of the year – the software is the basis for rec­om­men­da­tions on Amazon. Basically, there’s a trend towards more in-depth rec­om­men­da­tion systems in e-commerce. In addition to the pos­si­bil­i­ty of dis­play­ing 'Popular items', more and more companies rely on highly per­son­al­ized rec­om­men­da­tions. As a rule, several rec­om­men­da­tion strate­gies are brought together: pur­chas­ing interests, popular items, and other factors, such as product avail­abil­i­ty and price changes, are au­to­mat­i­cal­ly included.

In e-commerce, rec­om­men­da­tion systems are a par­tic­u­lar­ly important topic. This is because online stores have the pos­si­bil­i­ty to increase their con­ver­sion rate through suitable rec­om­men­da­tions and generate more sales. Many shop systems have in­te­grat­ed standard features for product rec­om­men­da­tions. This allows a solid analysis and cal­cu­la­tion– but the best way is to use a special software solution. Various providers offer companies SaaS (software-as-a-service) solutions. Well-known solutions in the US include Certona and Bar­il­liance. Most providers promise software solutions, which are in­di­vid­u­al­ly tailored and capable of self-learning, as rec­om­men­da­tion services based on their own per­son­al­iza­tion tech­nol­o­gy (model-based method). The great advantage of SaaS solutions is that they sig­nif­i­cant­ly reduce the time and effort required for im­ple­men­ta­tion. Store owners don’t have to invest in hardware or software. The mostly cloud-based solutions also feature a large range of functions. The software solutions take on three important steps: database tracking, feature en­gi­neer­ing, and data pro­cess­ing or analyzing at the end.

Tracking databases

To be able to analyze data, you have to collect it first. This can be done through classic tracking methods for most software solutions. Tracking includes relevant data on the location, shopping cart, date and time, behavior, and usually the com­plete­ly traceable customer journey. The program collects all this in­for­ma­tion and keeps it in a database.

Feature en­gi­neer­ing

When it comes to feature en­gi­neer­ing, the goal is to filter out features (i.e. char­ac­ter­is­tics, prop­er­ties) from the database. These features can be of a different nature, such as the time of the visit and its duration, the distances between actions, and many more. However, few features are of any relevance to the pre­dic­tions later on. The challenge that the system faces, is to precisely identify these important features. For this purpose, the system needs to find the char­ac­ter­is­tics that have a sig­nif­i­cant influence on the pur­chas­ing behavior, and ul­ti­mate­ly, the purchase decision. The in­di­vid­ual com­po­si­tion of the features varies according to the store, so an in­tel­li­gent analysis is required.

Pro­cess­ing and analyzing data

Based on the features defined for the online store i.e. the relevant char­ac­ter­is­tics and prop­er­ties, the system can now calculate pre­dic­tions for product rec­om­men­da­tions. Creating these prognosis models requires a huge amount of computing power, and sometimes takes several hours. The system saves the models, which then serve as a basis for cal­cu­lat­ing rec­om­men­da­tions. Every store visitor receives up-to-date tips and rec­om­men­da­tions tailored to them.

Per­son­al­iza­tion is becoming in­creas­ing­ly important in online marketing. This is not only due to the fact that companies feel under pressure from their com­pe­ti­tion and have to con­stant­ly strive to stand out from the crowd, but also because of users’ changed per­cep­tions. Today’s users are able to identify ad­ver­tise­ments more quickly and ignore them. If you are, however, able to attract attention by providing cus­tomized and relevant in­for­ma­tion, and by ad­dress­ing users directly, the chance of con­ver­sion is much higher. The same is true for rec­om­men­da­tion systems, which are becoming more sensitive and accurate. Finding the right strategy and reaching potential customers can have a positive impact on sales and the success of a company in e-commerce.