Ad Topic Hints

From Bitnami MediaWiki
Jump to navigation Jump to search

Facebook’s Ad Topic Hints enables marketers to engage people with interest attributes that people self-declare, which the authors call “person-driven advertising.”[1]


Ad Topic Hints proposal relies on a local client user profile containing attributes that can be used to improve the matching of content. The attributes are updated based on user interaction. The proposal suggests the interaction should be simple binary input (e.g., relevant / not relevant) rather than selection from a menu with too many choices.[2] The proposal suggests that the given the declared signal marketers would be more inclined to match more content associated with the user-provided hint.

The attributes are encoded into a “binary embedding vector” of a particular number of bytes. The proposal suggests the total bytes may be between 64 to 1024 bytes. The embedding is meant to classify rated ads into “topics,” belonging to an unspecified taxonomy. The authors acknowledge the common classification challenges and suggest that data sets to train the model should include “speakers of many languages and cultures, on ads for a wide variety of topics.“[3] Both the taxonomy and the embedding (classification) would need to be updated periodically.

The proposal suggests people could see all ads previously rated and modify their ratings. The proposals states that one reason to maintain the original ad and its prior classification is to enable reclassification into the new taxonomy or via the new embedding approach.

This proposal explicitly excludes from scope how sites match content to these attributes.[4] Yet Ad Topic Hints later suggests that people could indicate a preference to “explore [ads] outside their current choices.”[5] However, this “choice” violates the proposal scope not to define logic of the content matching systems. The proposal does raise the question whether preferences should be “always respected” or whether there should be a time decay or merely used as a weighting factor by content matching systems.

Ad Topic Hints notes the potential for bad actors to game the system with “click jacking,” or automating the feedback rather than ensuring it is user provided. To mitigate this risk, the authors propose visual feedback when interaction is provided, “You indicated the topic of this ad was relevant to you,” and then requires a second confirmation of this input: “Yes, correct” or “No, Cancel.”[6]

To diminish the likelihood of the array of Ad Topic Hints itself being used as a unique identifier, the proposal suggests combining user-generated hints with pre-populated hints as well as providing only a subset of the interest profile array during any given browsing session.

Ad Topic Hints suggests the scope of the storage should be at device level, rather user-agent, domain-level or identifiable user-level. This proposal does suggest people should also be able to download their array of topic preferences to port them from one platform to another. The mechanism to for interest portability remains undefined, especially if the storage systems do not allow for the authenticated match of the individual linked to this digital activity. Persistently storing digital activity linked to directly-identifiable identity would likely raise additional privacy concerns.


As with other user-controlled attribute proposals, Ad Topic Hints confuses the concern people have regarding receiving output content with the attribute data input that may or may not have been used to match the content. If someone would prefer not seeing running shoes, removing “sports” or “shoes” from their profile does not prevent matching on current context, geography or even registration data. To more directly address such user concerns, these proposals could be improved by storing sets of “paused” ads or brands, rather than unspecific list of topics as potential inputs that might lead to the unwanted content.

This proposal further confuses B2C content recommendations (optimized for user preference) vs B2B content optimization (optimized for marketer goals).

The authors may be optimistic about the utility people may derive advertising content matched to user-generated attributes, when they focus on portability of the user-generated topic hints. Given the browser is introducing noise and limiting the topic hints to subset of the interest profile array per session this will further diminish the likelihood that people will perceive the change in advertising content they receive, especially as all content matching logic is intentionally excluded from the scope.

Not all interest attributes are equally valuable to all marketers. By restricting the list of available attributes, or ranking them on criteria independent from publisher monetization, will further impair publisher revenues.

The proposal moves per-ad or organization attribute storage to the client, which will use greater client storage than needed at present. This forces people to absorb the cost of B2B storage and processing, that may require them to update older devices to continue to access ad-funded sites.

Open Questions

  • What taxonomy will be used and is it granular enough for marketers to derive sufficient value per the goal of proposal?
  • Is the topic taxonomy at a controller (e.g., marketer) level or product category (e.g., “consumer electronics”) level?
  • If at the product category level, how does it enable people to express preference (e.g., “I prefer Sony” and “I dislike Skyworth“)?
  • As there are millions of organizations, if each ad is stored locally what is the impact on local storage or classification process of interests?

See Also