Websites can be adapted to the situation in order to meet the current, individual needs of their visitors. As is often the case, an example helps to demonstrate the problem: one visitor to a web shop may be a "bargain hunter", while another one may prefer premium class items. If cheaper articles are displayed to the first customer than the second one, one can speak of a (partially) personalized web shop. What can be personalized? Personalization is not limited to displaying recommendations as outlined in the example above. In theory, every element of a web shop can be adapted to individual customers. Typical examples include
Following the working definition above, personalization is the situational adaptation of a website. A situation is defined by a list of features, and the website is then adapted according to the feature values. Features that can be personalized can be divided into the following three categories.
The difference between the first and the other two categories is based on their availability: While personal interests only become apparent over the course of the customer's interaction with the website, the information listed under 2 and 3 becomes available within milliseconds of the page visit and can be used for personalization from the start. In case of returning visitors, past personal interests are also available from the beginning of the session.
It's very simple: Customizing the user experience has been proven to have a positive effect on purchasing behavior.
The multitude of customizable elements on the one hand and the variety of diverging customer needs on the other hand can result in an virtually unmanageable number of possible personalization options. A successful strategy amounts to knowing which of the alternatives are the optimal ones for any specific customer. Another complication arises from changes in the user behavior over time. Established rules according to which web shops are personalized can lose their validity. Humans quickly reach their limits in finding and maintaining this set of rules. As a result, effective personalization requires a form of automatism. This situation calls for a special form of Artificial Intelligence: so-called self-learning software, i.e., software that can independently find out how the elements of a website have to be adapted to meet the needs of the customer.
A precise definition of the goals of personalization is required. The user selects the KPIs that are to be optimized by adapting the elements of the web shop. This adaptation is usually subject to business logic, i.e. boundary conditions that define the scope of actions of the self-learning software.
With self-learning software, a so-called artificial agent learns from the results of its interactions. In the context of personalization, this means the following. The artificial agent adjusts an element of the web shop for a user and registers the resulting behavior. If a change to the product description resulted in the purchase of a product, taking this measure was probably beneficial. Given the current situation, it is consequently "upgraded" to a certain extent. Formally, this means that the corresponding parameters of the algorithm are adapted. The artificial agent then learned the following: This or a visitor with comparable characteristics responds positively to the updated product description. If the same visitor returns at a later date, or for a comparable visitor, this measure should be repeated in most cases. But why only in most cases, why not always?
User behavior changes over time, and with it the optimal personalization strategy. Apparently, users are also changing their behavior at an increasing pace. This may be due to the improving transparency, such as the availability of comparison portals. For example, those who are unable to lower prices quickly enough are prone to losing customers who have found the product they are looking for cheaper elsewhere.
Paradoxically, an Artificial Intelligence that truly deserves its name has to act less intelligently from time to time. So it has to try something less optimal, like adapting a product description that has so far worked less well than another. But why does it have to do that? The answer is simple: this product description may actually work better. Maybe because it "always" did, but maybe also because visitors have changed their behavior in the meantime. In addition to taking advantage of existing positive experience, an artificial agent should also always invest a little in the exploration of supposedly negative experience. Like us humans, Artificial Intelligence can learn from trial and error and thereby adapt. While people are expected to adhere to the rules of society, the actions of an Artificial Intelligence may be subject to business logic. That is, not every trial is allowed.
In both cases (exploitation or exploration), however, it is important not to make the same mistakes over and over again. It is therefore desirable to have a concept of comparability of situations as introduced below.
The previous section talked about comparable situations or customer behavior. As mentioned, having a concept of comparability is an advantage. The reason for this is that with the multitude of possible situations and the diversity of customer behavior, even with a wealth of experience, there will always be situations that are different from those seen up to then. Anyone who can then make comparisons to what they have seen before, or who can generalize previous experiences, will be able to make an informed decision despite the new situation.
Formally, this means identifying unchangeable features of situations that require the same optimal decision. In this context, we also speak of pattern recognition. In the customer's example, it may be that the effectiveness of a particular discount campaign, for example, does not depend on the day of the week, but does depend on whether a public holiday is imminent. An upcoming public holiday would therefore be one of these unchangeable characteristics sought after.
Deep learning, another method of Artificial Intelligence, has the outstanding property of independently identifying those features in data that allow for the best possible generalization. As a result, in some fields of application, such as the diagnosis of leukemia, a performance has been achieved that is on a par with humans. These algorithms were often able to detect and exploit features that were not known to a human expert. As a result, deep learning is used, for example, in the design of medication, the recognition of language, or the restoration of images. The combination with reinforcement learning allows users to enjoy the advantages of both worlds. Appropriate algorithms make it possible to make sustainable decisions even in unknown situations by means of generalization. However, deep learning is not always the best choice. To achieve good performance, it requires large amounts of data for training. Not every application complies with this requirement. That is why we do not rely on a single algorithm, but test a number of complementary algorithms that can take the helm depending on the current situation.
In summary, it can be said that the effective personalization of websites requires two basic skills. 1. to generalize and 2. to be able to adapt. Common software services are moderately successful in these tasks.
A crucial flaw unites all existing solutions, however: elements of websites are personalized independently of one another. They unknowingly or against better judgement neglect the fact that all personalized elements together make up a joint user experience. This is often sold as an advantage: a long list of customizable elements is presented as a complementary range of services for marketing purposes. Product recommendations are used, for example, independently of personalized discounts. However, it is axiomatic that a customer has to find the right product at the right price, otherwise they just won't buy it.
The left hand literally doesn't know what the right hand is doing. The user experience becomes a patchwork of separate optimization attempts, similar to a poorly managed company, whose departments optimize according to their own key figures and there is no coordination with regard to the overarching corporate goals. The result: Potentially significant parts of the profit are not realized. Another example should help: In football, a well-rehearsed team of only average players can beat a team of gifted but selfish soloists. Also in e-commerce, a harmonious interplay of the personalization of all elements is required.
Since all these hopefully convincing arguments are in favor of controlling the overall personalization effort of all elements by a single AI, the following paragraph is intended to show how this can be implemented in practice.
In practice, this means that in every situation all customizable elements and all their variations are presented to the agent for selection. The crucial point here is that this list is not artificially restricted to individual elements of the web shop, but includes all combinations of possible variants. Accordingly, it contains, for example, options for changing the layout and design of the webshop, a list of possible product recommendations or discount options. With two product variations (blue and red) and two possible discounts (5% and 10%), this implies four different combinations (blue-5%, blue-10%, red-5% and red-10%) that the artificial agent can choose from. By monitoring the resulting user behavior, the agent can then identify particularly lucrative combinations in a user-dependent, i.e., personalized, manner.
State of the art personalization therefore requires Artificial Intelligence, which, in addition to good generalization properties and autonomy in interacting with users, can deal with an overwhelmingly large number of possible options. Contrary to conventional approaches, a suitable artificial agent is set no intrinsic limits. If there are a sufficient number of interactions with known user behavior available to learn from, it will always be able to derive effective recommendations for optimal action selection.