|
Using concepts for Ad Targeting - Click here for original Patent filing
This patent deals with serving Ads based upon keywords or moreover, topical ‘concepts’ to help and determine ad ‘relevancy’. Yes, here I am once again discussing ‘relevance’ – I hope 2008 brings me a new hobby. This time I am looking ( thanks Bill ) at the Ad serving world. I do feel it is important as many of the concepts relating to establishing meaning or ‘themes’ to content/information/words is much the same in this area is it is in the ‘organic’ search processes.
There is a bunch of the usual mind numbers as far as the nuts and bolts of user account setting such as geo-location, budget, keywords desired, to the server side calls for number of ads to be displayed to size/real estate available for them. Of interest there is the checking for geo-location issues from say a search query to a regionally targeted ad (sounds like IP delivery er.. cloaking to me?).
What’s the point?
In describing the desire there is an interesting line;
“an advertiser may attempt to target its ads to more narrow niche audiences, thereby increasing the likelihood of a positive response by the audience”
And that’s is the crux of it; trying to define the Ads and the target audience as tightly as possible. By doing so the conversion rates should improve and (in theory) the advertiser can get more bang for their buck. Also, in a Pay-Per-Click model, the more targeted and higher quality the Ad, the more chance of getting the ‘click’ that the Ad server wants as per the business model. Tighter relevance and higher quality ads are an obvious benefit and motivation.
The parameters are defined by the advertiser at the time registering the ad who establishes, “one or more concepts to be used to target its ad, or indicate whether some concept indicator is relevant to its ad.” So how is dinner being served?
“candidate ads (that) have been determined to be relevant to the request using, at least, keyword targeting information.”
If you already understand some of the concepts used in indexing and scoring of documents based on relevance in organic search indexes, much of the usual concepts are the same as far as keyword/phrase and the processing of ‘concepts’ is concerned. Of interest is that the ever popular ‘jaguar’ example is back (used in some recent relevance patents);
|
“As another example, the query term "jaguar" could refer to the car by that name, the animal by that name, the NFL football team by that name, etc. If the user is interested in the animal, then the user might not be interested in search results which pertain to the car or NFL football team. Similarly, the user might not be interested in advertisements, targeted to the keyword "Jaguar," but that pertain to the car or NFL football team.” |
So we are still looking at words as phrases or ideas/concepts as opposed to simplistic keyword models – which is termed as ‘concept similarity scores’. Most of the tightening or relevance pertains to this train of thought. See earlier ‘phrase based indexing and retrieval ‘ to get some ideas on these concepts or documentation on things such as ‘probabilistic latent semantic analysis’ - Latent Dirichlet Allocation - as they are a good core towards understand relevance in the indexing and retrieval process.
Also of interest in this method is that it;
|
“tracks such performance and uses it to modify concept scores. Concept performance information management operations accept the performance of concepts in ad serving and may adjust concept performance information accordingly. The concept performance information may include a number of entries, each including a concept and at least one performance factor (such as a weight for example).
A performance factor may be tracked for one or more of (a) a concept source, (b) a concept in general, and (c) a specific keyword-concept relationship. Thus, for example, if an ad is served pursuant to a concept, from a concept source, because of the concept's association with a request keyword, one or more performance indicators of the ad (e.g., click-through, conversion, etc.) may be tracked and used to adjust a performance factor(s) of one or more of (a) the source of the concept (e.g., ODP, a classification technique such as a semantic classification technique for example, etc.), (b) a concept in general (e.g., across all source and/or all keywords), and (c) a keyword-concept relationship (to reflect the fact that the same concept may perform well when used for ad serving based on its association with one keyword, but may perform poorly for another keyword).” |
So, much like some other works lately, ‘performance’ or what some would call ‘click data’ or ‘bounce rates’ and so forth, come into play in the endeavour to find relevance in words/concepts. What people are doing is as important to determine intent as understanding the concepts of a give document/ad text. I can’t help but think some form of communication could not be somewhere to pass scoring elements from one ranking process to another.
Scoring Elements
The method discusses taking a core set of candidate documents that are determined by ‘(at least) keyword targeting information’ as well as further scoring is also achieved via;
- Ad performance information
- advertiser quality information
- ad price information
- an information retrieval score.
In many ways it is not only using the data to decide which Ads are to be placed in a given setting, but the order (ranking) of them as well assigning a ‘relative preference attribute’. The content of the target page is established via concept with a ‘concept vector’ attribute. Delivery of the Ads is based on understanding the delivery page concept and assigning Ads based from the similar concept attributes mentioned earlier.
Using past data for scoring
The scoring/relevance of the Ad, ( as per this method ) is also targeted/scored for relevance including such factors such as tracked performance information. This area can include previous processed search queries to which the Ad would have been relevant. Targeting the ‘concepts’ can also include “other ads using the same or similar targeting criteria information”. So past runs of existing ads and comparison to other ads can come into play. The performance of past Ads that have been served that are related to the ‘concept’ can be used as further scoring of the current set(s) of results to be scored.
This information can include advertisement selection/conversion data. The score can be increased or decreased based upon the performance being above or below a pre-defined threshold (performance level). I must say they are more than a bit hazy on what a ‘conversion’ is, they have some variants – the most likely being;
|
“The ratio of the number of conversions to the number of impressions of the ad (i.e., the number of times an ad is displayed) is commonly referred to as the conversion rate. If a conversion is defined to be able to occur within a predetermined time since the serving of an ad, one possible definition of the conversion rate might only consider ads that have been served more than the predetermined time in the past.” |
One can see how such scoring can also be used when deciding if a set of Ads appears in a search result page as opposed to a content ad placement (website). They can use relevant ads that have a higher historical click/conversion rate, to capture ‘clicks’ at a higher rate (conversion), in an advantageous location such as high traffic search segments (travel etc..). By increasing the likelihood of a conversion, they increase the revenue model. That’s what I would be thinking at least if it was my toy.
Less historically desirable or under-performing Ads, could be designated for a lower class of traffic locations. Sure, that could be a bit in the ‘conspiracy theory’ terrain, but it most certainly could be modelled that way.
Summary -
In simplest terms, much of the scoring is based upon;
(a) ad performance (b) ad price information, (c) advertiser quality (d) IR score
As always, we’re not setting the thresholds for each, so how ads are scored/ranked can vary. I can see how it certainly encourages a well written, well placed ($$) ad. If no one clicks your ad, they don’t make as much and the more they make on your ad (ad price) has its own metric. It’s only natural, oui? There is some more about the ‘similarity score’ but I don’t want to go to far into that with an older document as more recent methods may superseded to at least have had an effect on this particular process.
Published methods since this was filed seem more suited as far as ascertaining relevance. I would have to imagine that this system would have benefited from being able to attain a ‘similarity’ score from some of these other methods/models – or other merging. The underlying concepts derived from end user data as far as trying to achieve further ‘relevance’ further brings to light the use of such data as a scoring set. Also of interest is the combination of similarity scoring with ad performance scoring variables – something likely not commonly considered.
All Ads were not created equal.
(also see related parts; Part I, Part II and Part III and the summary on my Blog)
Resources - this is part of a 3 part series - Summary; Relevance through end user metrics - Learning a probabilistic generative model for text - Ranking documents based upon large data sets - Using concepts for Ad Targeting.
Original Patent - Ranking documents based on large data sets
Patents of further interest - Predicting ad quality - Using estimated ad qualities for ad filtering, ranking and promotion - Using estimated ad qualities for ad filtering, ranking and promotion -
|
