This is the summary of an article by Phil Tian, et al. which talks about a proposed combining Anonymous Video Analytics (AVA) and Data Mining techniques in what is called as Intelligent Advertising Framework (IAF). You can get the pdf of the behavioral targeting article here: Realizing Targeted Advertising in Digital Signage with AVA and Data Mining.
Digital Signage is described as using projected displays in public places to display advertisements, news, and similar announcements. Digital Signage has the second highest revenue growth in advertising, next to internet advertising.
When it comes to targeted advertising, that concept is still new for digital signage. Targeted digital signage could mean advertisements shown in public places will not be about random things, but would reflect the traits of the audience looking at it.
The study proposes an IAF, or Intelligent Advertising Framework, that combines two techniques to achieve targeted digital signage. These two techniques are Data Mining technologies and Anonymous Video Analytics (AVA).
Architecture of IAF
The main components of the IAF are the analytic server, data mining server, data mining module (DMM), content management system (CMS), AIM Suite, and Digital Player.
AIM Suite is a component of AVA. It analyzes the audience video feed and sends viewership information to the Analytic Server, which then cleanses this information and saves it to the Cloud. The Data Mining Server is responsible for creating advertising models. DMM is responsible for a lot of tasks, including model query, predictive modeling, data connection, and rule extraction, and making these processes available for other components. What DMM learns is sent to the CMS, which uses other ad information as well to create lists of customized advertising. The Digital Player makes the final, real-time decision as to what ads to display on the digital signage. An ad playlist is also created and sent to the Cloud’s data repository.
Anonymous Video Analytics
AVA is based on computer vision theory. It has three major steps, human face detection, demographics recognition and viewing event creation.
In human face detection, an optical sensor that is located in the panel captures video feed, and face patterns are analyzed by the AVA. Demographics recognition is done when a machine algorithm learns to analyze certain pixel combinations and connect them with age and gender, among other demographics. Finally, in viewing event creation, AVA detects a viewer and determines when he/she started and stopped viewing the advertisement and for how long.
Several statistics and reports can be obtained from this information, including average attention span and total number of viewers. Furthermore, the system is privacy-friendly; it doesn’t identify the names of the faces.
Data Mining for Targeted Advertising
Data Mining Technology is used to discover viewing behavior patterns of the digital signage audience. The steps involved include multiple advertising model training, audience targeting methods, weighted audience counting, passer prediction models, and ad selection based on Advertising models.
In multiple advertising model training, the models is trained regularly and sometimes on-demand when it is showing a sub-par performance or if requested by operators or users. Audience targeting methods include seeing based targeting, prediction based targeting, and context based targeting. Weighted audience counting is done for passers based on the time points as they pass the digital sign. There are two passer prediction model types used: passer distribution prediction model and dominant passer prediction model. Finally, ad selection is based on both seeing based targeting and prediction based targeting.
Targeted Advertising Process
The targeted advertising process is divided into three parts: learn advertisement models, create default playlists and finalize and play the playlists.
The advertising model attempts to recognize patterns, including what audience group is interested in what kinds of ads, to what extent do this audience group exhibit such interest, and where, when, or even under what weather types of questions.
The models are then transferred to the CMS, which in turn creates the list of Ad Categories. After some shuffling these ad lists are then transferred to the digital player, which operates both in offline mode or online mode. Ads obtained from offline mode are based on from default playlists based on time schedules while online ads are calculated from advertising models. The digital player switches between both modes depending on the advertising models’ confidence level.
