Ever been confronted and annoyed with past online search behaviour in your recommendations, in for example advertising? I personally have been getting Facebook banner advertisements of the “PROS Price Optimizer™” software since our Digital Transformation Project for the BAC course entailed implementing price optimizing technology at hairdressers. Obviously I’m not really planning to buy this software, but it represents how past search and click behaviour is the basis for our current recommendations. In addition, it shows how inaccurate it can be and that past search behaviour is not always representative of our current wishes and can lead to some dissatisfaction about these recommendations.
In their article Timing of Adaptive Web Personalization and Its Effects on Online Consumer Behaviour, Ho, Bodoff and Tam (2011) describe how they studied the effects of recommendations based on either previous or current behavior. They make a distinction between static and dynamic advertising: two similar forms of web personalization, yet different in the ‘timing’ dimension. Static personalization uses only previously known facts about the consumer, such as demographics and all the preferences collected based on search history. Adaptive personalization on the other hand, uses choices made in one particular search session.
I want to use this blog post to show the important findings of this study instead of focusing on summarizing the entire article. As I’m writing my thesis on recommendation agents (RAs), I am usually very interested in reading background theories and models, but I would not recommend anyone to read this entire article. My opinion is that the authors were a little too enthusiastic in the theoretical part. They got too deep into the mathemetical models of the consumer search theory and spent too little words on the actual findings and implications. So my advice to the interested readers: skip the first half of the article!
That being said, I would like to show the differences between the two mentioned personalization types by looking at my own YouTube recommendations. I opened the YouTube homepage today and, like always, got recommendations based on my previous clicks that are recorded on my gmail account.
Figure 1 – “Early” recommendations
Youtube knows I like Adele, which is why I get two recommendations for that. I also have a slighter worse taste in music and like to listen to the songs I liked when I was 8, which is probably why I’m getting a recommendation for K3. I’m also seeing a recommendation for a Britain’s Got Talent video, which I think is something I would like right now, so I click that video.
These types of recommendations are the ‘early’ recommendations, because YouTube has to base this on my previous search behaviour and has zero click stream data from my current session. Websites might choose to alter these recommendations once my session is providing them with more information, or they might keep the recommendations the same. If I would get the same recommendations after spending for example half an hour on YouTube, then the form of static personalization is used. YouTube however uses adaptive personalization, so I know the recommendations are going to change as I browse along.
So what happens when I’ve watched the Britain’s Got Talent video? My recommendations turn out to be slightly different. The video I just watched is replaced by a video of Dancing With The Stars.
Figure 2 – Adaptive personalized recommendations after watching 1 video
YouTube is probably thinking that I might want to watch all kind of talent shows, but in fact I’m not really into the dancing scene. Imagine I’m looking for a good singing audition to show my grandmother. So I search for Britains Got Talent videos in which people sing instead of dance. This way I’m giving the recommendation agent somewhat more info about my current interests. After browsing some more, the following recommendations are given:
Figure 3 – Adaptive personalized recommendations after watching 3 videos
Figure 4 – Adaptive personalized recommendations after watching 5 videos
The recommendation videos in the pictures show that once my browsing session is happening, I’m getting more accurate recommendations based on only this session. Adele is still present at some point because the RA probably isn’t sure yet, but after a while he full on goes for the singing audition videos.
So what does this have to do with the article? What I am trying to show with these pictures is that there is a tradeoff between quality of the recommendation and the marginal benefit of adding an extra product to my current choice set. Indeed, after a while the recommendations are pretty accurate, but do I still really need them at that point? I have probably found a good video to show my grandma already and don’t feel like watching any other video anymore. So this is where timing comes in. The longer an RA waits with giving recommendations, the more accurate they’ll be but the less value they will have for the consumer.
The most interesting part of this paper in my opinion is that they also tested satisfaction among these different types of personalization. The study indeed showed that consumers confirmed that adaptive personalization provided them with more accurate recommendations when the RA waited longer, yet the satisfaction about the RA also dropped when recommendations were received ‘later’. This might all seem pretty straightforward, but it is very important to keep in mind for websites. Managers need to decide whether satisfaction or accuracy is more important for them and need to find the optimal point in timing their recommendations. The article does not give a solution for this, as this optimal point differs for each context and product type. On that note I would like to say that the article describes some logical findings and provides a good attribution for RA research in satisfaction.
Figure 5 – User satisfaction decreases the later the RA gives recommendations (Ho et al., 2011)
For companies like YouTube and Facebook, these kind of issues have already been tackled sufficiently. Adaptive personalization has been optimized in a way that there are multiple timing points for changing recommendations, yet static recommendations are also still incorporated. However not all websites have an advanced IT team for this, or they might not have the contextual opportunity to provide this many recommendations.
For future research I would like to know more about when you know a customer is not happy with the recommendations he is getting. For example, if I haven’t clicked on the PROS advertisment after the 56th time I’ve seen it and I haven’t searched for price optimizing software anymore, there must be some algorithm be able to tell Facebook that I’m not interested in PROS at all? (I know I can tell Facebook I’m not, but I refuse to do that.) These are things research is fortunately heading towards now and I’m looking forward to what future studies will be able to tell us.
Ho, S. Y., Bodoff, D., & Tam, K. Y. (2011). Timing of adaptive web personalization and its effects on online consumer behavior. Information Systems Research, 22(3), 660-679.