The Recommendation Technology we developed in-house is at the heart of plista. Through the interplay of dozens of algorithms, it enables extremely precise targeting with high hit rate and low scatter loss. This way, the delivery of advertising and content can always be matched to the individual preferences of a user. Advertisers, publishers and users alike benefit from the high relevance of plista recommendations.
Take a look behind the scenes of the plista Recommendation Engine und get to know the key technologies responsible for the accuracy of our recommendations.
Data is the basis for plista's precise user-individual recommendations. Therefore, we process huge amounts of it. Each day plista servers record about 600 million queries, also called events. An event arises each time a user visits a page having an implemented plista widget.
Events are every bit as unique as the user who triggers them. plista captures more than 100 different attributes per event in order to be able to exactly define the context of each event.
It's primarily about “hard” attributes, such as the browser and browser language, operating system, device and model, date and time, geolocation, and much more. Add to that information about previous websites visited, article categories (such as news, sports, business, fashion, etc.), semantic information about the content just read, and publisher ranking within the lists maintained by AGOF, Comscore or Nielsen, etc.
Data collection follows the ETL principle (extract, transform, load). It assures the anonymization of the user, among other things. After extracting the data from the different sources, it is transformed in order to make it usable, a process which includes anonymization even before storage. From that point on, the user exists in the system only as a number, thus preventing any link to the actual person. Moreover, data such as name, address, etc. are not even collected.
It is on the basis of this collected and organized data that the plista Recommendation Engine makes customized reading and advertisement recommendations for each user. The process uses several dozen algorithms for which the plista Ensemble Algorithm acts as overall conductor, orchestrating them in real time. The most important technologies used include:
Behavioral Targeting allows the derivation of users' behavioral patterns and thus their interests from the collected data. Visited websites, previous interactions with ads, online purchases, etc. help to gather user preferences and to bring the user into the process in order to be able subsequently to present individually customized recommendations. For example, should a user visit travel websites or read articles about travel, one can assume he would also be interested in trips. Based on that, he would receive ad recommendations about vacation trips, airline offers or tourism-related articles about, etc. During this process, behavioral targeting takes into account both the past and current interests (as seen in the browser history as well as the website currently being visited).
The key technology used by the plista Recommendation Engine is Cross-Domain Collaborative Filtering. This highly effective use of Behavioral Targeting is based upon the assumption that users having interests in common also take an interest in similar subjects. To this end, the user's reading and click behavior are not only evaluated, but also compared with one another. Groups are formed whose users demonstrate similar behavior patterns and interests. These are identified as statistical twins. For example, if one of the twins is interested in the subject of real estate, the chances are good that the other twin is likewise interested in it. And accordingly, he receives similar recommendations for content or advertising. The more data available about the users' interests, the better Collaborative Filtering works.
In order to deliver the best possible results, the plista Recommendation Engine also takes into account the editorial environment in which the recommendations are embedded. For this purpose, semantic analysis is used for much more than only searching for individual keywords. The plista Recommendation Engine performs a comprehensive content analysis of the respective page, during which the content is precisely categorized, for example as “news,” “sports,” “tennis,” “Wimbledon,” etc. Based on that and using more extensive text analyses, semantic similarities with other articles and advertising from the plista database system can be identified. In that way, plista recommendations can be easily assigned to articles covering appropriate subjects: A sports article advertisement or a report about the current Wimbledon winner will thus be displayed under another tennis or sports article. This increases the subject relevance of the recommendations and thus also the user acceptance.
Depending upon optimization goals (CTR, view time, conversions, etc.), certain recommendations are deemed especially suitable. plista links the performance of the recommendations to the different attributes and creates a list of top recommendations for every attribute (publisher, device model, browser, etc.). For example, article D attracts the most clicks among iPhone 6 users, and article X gives the best results for users located in Berlin, etc.
In order to generate recommendations that are perfectly matched to the target, the plista Recommendation Engine combines together the most popular lists of several attributes in real time. The more attributes that are consulted with their lists, the more personalized the recommendations will be.
This approach is excellently suited to avoid „cold starts” - that is when there is little or no information about a user. This usually occurs when the user triggers a plista event for the first time or has deleted cached cookies.
Besides the core technologies described above, plista uses a variety of other algorithms which combine together the most diverse parameters. That is how some of the algorithms optimize delivery of articles based on topicality by taking into account such things as the age of the article at which the reader's interest begins to decline, for which type of publisher the topicality is more important (such as a news website vs. a special interest website), and also the user's individual interests for topical content.
By virtue of its wide spectrum of powerful recommendation approaches, the plista Recommendation Engine is in a position to always offer the best solution.
The plista Recommendation Engine comprises dozens of algorithms that simultaneously process a large variety of parameters in real time. The plista Ensemble Algorithm takes charge of coordinating the processing complexity, acting as conductor to achieve perfect harmony within the algorithm orchestra. It decides based on the individual context of an event which algorithm ought to be deployed to achieve the envisioned goal (such as generating clicks etc.). Usually, the best solution is to use a mix of several algorithms: Combining several procedures can compensate for any drawbacks and generate for every scenario the best customized recommendation for the situation at hand.
But to do so, the conductor must first find out which algorithm works best, so several are tried against each other in real time. The conductor assigns weights to their influence in the final outcome based on their performance: The algorithms that perform well are weighted higher and vice versa. Of special note here: Algorithms that perform poorly are given a chance to improve, since we know that their potential is sometimes only demonstrated over the long term. That way, meta learning becomes possible – a continuous learning process which improves both the plista Recommendation Engine as a whole as well as the results for advertiser and publisher.
The meta algorithm is also able to pursue several goals simultaneously (click and engagement rates, time spent on site, conversions, CPC, etc.) and provide the best balance.
Processing the data in real time plays a key role for the quality of the recommendations. The goal common to all algorithms is thus to make available as knowledge and as quickly as possible the gathered information and/or user feedback. As soon as a new preference is gathered, it immediately enters into the next recommendations in order to improve their precision.
Scroll To Top