This is the summary of the article by Koen De Bock and Dirk Van den Poel, which presents a model for analyzing demographic information and click stream data to create web user demographic profiles. You can get the PDF of the behavioral targeting article here: Predicting web site audience demographics..
The online advertisement industry is a very large industry, and is growing rapidly. Thus, many advertisers and researchers are motivated to improve the effectiveness of online ads further. One of these methods is advertisement personalization.
Advertisement Personalization
This method aims to precisely target advertisements to online users based on their web characteristics. An example is behavioral targeting, which tracks online information of users using clickstream data, search term usage, among others. In addition, behavioral data can be combined with user specified preferences, customization settings and demographic information to further enhance its capabilities.
Demographic information collection
Originally, demographic information is used as a targeting tool for old media advertising efforts, and is deemed inferior to behavioral targeting. However, demographic is shown to be the second most important option after behavioral targeting according. This is true because brand building is one web advertising function that can benefit from this information. Brand building doesn’t have to have direct response as a result of interactivity, which is beneficial for behavioral targeting. Online, collecting demographic information is challenging because of the anonymous nature of web activity. However, there are several solutions for this, including user registration, and having demographic profiles through web surveys. There are several issues to these solutions however, such as costly efforts, visitor annoyance risks and applicability to only a few websites.
Clickstream patterns plus Demographic Attributes
This paper proposes a cheap and effective way to create a demographic profile of web users. This involves collecting click stream patterns and demographic attributes and applying the two to Random Forest classifiers to create demographic attributes of anonymous visitors to a particular website. Organizations which use this method can benefit by using the results as additional data for behavioral targeting, or as demographic predictions of their visitors.
Model Training and Scoring
The first phase in the methodology of this study is the model training phase. Here, data is collected to train the Random forest predictive models. These data are demographic info from random online surveys, and clickstreams from server logs. These two then serve as input for training the models. The next phase is the scoring phase, which is the application of the Random forest classifier models. In this phase, demographic profiles are obtained. Unlike the previous phase, this phase can be done repeatedly for various websites.
Data Collection and Model Verification
The model validation and demonstration is done through a Belgian organization. Data was collected twice, in September 2006 and February 2007. The respondents were randomly surveyed to obtain demographic info, and cookie tracking was used to collect click stream data.
Results
Results in this study show that the Random Forest models are superior compared to other benchmark algorithms. Furthermore, these models are most accurate in predicting gender, which is accurate at 69%. The other three demographics are under 50% accurate, but still do extensively better than other algorithms. On average, the obtained demographic profiles in this study are good, showing that this model can be used for applications in business, helping managers choose the websites they will use for online advertising. This also shows that demographic information is still relevant for web advertising.