Reinforcement Learning in Social Media Marketing

Reinforcement Learning in Social Media Marketing

Patrik Eklund (Umea University, Sweden)
DOI: 10.4018/978-1-7998-5077-9.ch003
OnDemand PDF Download:
No Current Special Offers


In this chapter, the authors describe an architecture for reinforcement learning in social media marketing. The rule bases used for action selection within the architecture build upon many-valued (fuzzy) logic. Action evaluation and internal learning is based on neural network like structures. In using variables measuring the effect of advertising, we must understand direction of influence between advertiser, owning the content of the advertisement, and advertisee, as the target of an advertisement, and as facilitated by social media marketing. Examples are drawn from Facebook marketing.
Chapter Preview


An Ad conveys a message, and is represented by keywords and the content of the Ad, containing text, images and video, annotated with links and objects (like buttons) enabling and inviting the Advertisee to interact with the Ad, and become influenced by the Ad. An Ad is like an expression presented (syntactically), and perceived (semantically) by the Advertisee. That perception depends on the Advertisee profile and behaviour. Ads with video content are either digital video Ads or as based on programmatic video advertising.

The content of that Ad can be dynamically modified, and the characteristics of the Ad should be described as fine-granularly as possible. A coarse-granular description of the Ad leaves less room for subjecting the Ad itself to modification during optimization of the campaign.

Ads are part of Ad sets, and a set of Ad sets constitutes a Campaign. This means that Advertisers must deal not just with the fine-granular description of respective Ads, but also consider the characteristics of the Ad sets, and the structure of the Campaign.

Facebook generates cost related to Ads, so that costs of Ad sets are aggregated costs. A campaign cost is the accumulation of all costs incurred by specific Ads and Ad sets.

Bidding and Facebook’s Ads auction is not only price related, i.e., the highest monetary bid does not necessarily win, but the winner is the Ad that creates the most overall value. The mechanism for calculating that “overall value” is hidden. Apart from the Advertiser Bid, being manual or automatic, Ad quality and relevance comes into play. An auction takes place whenever someone is eligible to see an Ad. Manual bidding influences reinforcement learning differently as compared to automatic bidding.

Ad performance can be measured e.g. by engagement and brand lift. Engagement with the Ad, can be measured either by hovering, scrolling or interaction with an extension, scroll velocity as a proxy for attention, how long a native Ad is viewed (even if not clicked), average time reading or watching, and so on. Brand lift can be measure e.g. by shares, followings, and email subscriptions.

There is a continuously on-going debate about performance metrics, but the What of measuring seems not properly connected with the Why of measuring. Metrics has been focused on outcome and performance, but the trend is to focus more also on Ad content. Formal terminology and nomenclatures, and as based on formal logic [Eklund et al 2014], for classifying Ad content is largely missing in marketing.

Performance metrics include

  • Actions

  • Clicks

  • CTR

  • Impressions

  • Relevance score

Key Terms in this Chapter

Reinforcement: An unsupervised learning algorithm.

Social Media Marketing: Marketing using media channels provided by social media like Facebook, Instagram, etc.

Neural Network: A supervised learning algorithm, based on layers of weighted sums, suitable for classification.

Fuzzy Logic: A logic language involved many-valued truth.

Complete Chapter List

Search this Book: