Guest post by Paul Golding, Full Stack Engineer and AI Specialist at Meta
With recent advances in AI towards language understanding (LU), there is increased enthusiasm for digital marketing applications. I recently explored an application in personalization of financial services. Initially this was to alter the “voice” of a chatbot, but also to modify marketing messages (e.g. emails).
Why make Personalization more Intelligent?
Personalization, the holy grail of digital experiences, is easier said than done. For infrequently used services, like financial applications, we need an innovative approach that mitigates the problem of not knowing much about the customer (unlike e-commerce).
I was motivated by previous work in art recommendation by the use of personality types to tailor marketing messages. For example, someone identified as adventurous might receive messages with adventure-biased language and images. I was also motivated by research that suggested a correlation between spending habits and personality type.
A Workable Foundation
I will superficially describe the outline of the research as a taster for what might be possible. I’ll leave technical details for another discussion.
Personality profiling is possible using the 5-factor model. Users can be asked to fill-out personality tests online to provide a training population. If those users allow access to their spending profiles (e.g. via bank or credit-card spending) then we can use AI to predict personality profiles based upon financial records. This assumes access to such data. In Fintech environments, this is typically facilitated using banking APIs like Plaid. Numerous Fintech services are built upon such APIs, like the many “finance-buddy” or savings apps.
This is a simplistic outline, but a workable foundation for correlating financial transactions to personality type. Some of the details I’m skipping over relate to finding a more succinct method of ascertaining personality. Five-factor profiling typically requires a burdensome number of questions (called “instruments”). If a large-instrument test is conducted initially, then machine learning can be used to weight instruments in relation to the available discriminatory information in the financial dataset. There’s no point in digging deeper into more nuanced personality aspects if the financial data doesn’t support such discrimination.
An additional consideration is how to extract a set of more meaningful questions (in terms of spending habits) that might support a kind of “quiz-like” user experience such as Molton Brown’s Fragrance Finder. In my research, I explored the formulation of questions that had the overt purpose of tailoring a financial solution. It also had the covert purpose of identifying personality type. This would allow rapid personality assessment to be built into the onboarding experience for all users, including those with little or no financial data available (e.g. prospective customers).
But the question still remains as how to tailor marketing messages using personality type. Here we can deploy a number of tactics. Firstly, we return to our profiled population and ask them to rank a variety of phrases (and images) using a scoring method. This provides data to create a language model that is personality-biased. There are important details here in the set-up. For example, using training data sampled from corpi more replete with metaphors and poetic descriptions, can amplify the potency of the process.
In the first instance, it seemed easier and more practical to build a machine that might assist a professional copy-writer in evaluating creative in terms of personality. As is typical with the application of AI, one has to evaluate how practical and useful the various levels of automation might be. Using a machine to score prose by personality profile is one level of implementation. The next is to use the language model to make suggestions. Think of a grammar-checker app that acts like a “personality checker” app instead. We could implement this in a number of ways including synonym suggestion, phrase suggestion or phrase exemplar suggestion. This could include a library of phrases that serve as examples that resonate with the target personality.
I have found that many pre-built language models (e.g. word embeddings trained on public corpi) are often too noisy in terms of synonym suggestion. In other projects, I have had much more success with training my own synonym library boosted by curated synonyms found in APIs like the Oxford English Dictionary.
We could consider what the outputs might be from this process. Let’s assume we know that a customer is interested in a loan for a vacation. Users identified with the adventure trait might see creative like: “Seize the day: book your dream vacation now!” Whereas, users identified with a simplicity trait might prefer: “Find your inner Zen by taking a vacation.”
More to Consider
Additional considerations that I explored were how to use personality to discover other user preferences. These could include when best to email someone and how best to formulate a subject line. Indeed, given that an email campaign is likely to involve a series of emails, it is also possible to construct in a way that identifies personality type. It then adapts accordingly on a per-cohort basis rather than an entire population basis (as many A/B tests do).
Also, the specter of bias is an important consideration, especially when dealing with financial services where regulations are well established for issues like disparate impact. As a foundation for broader use of AI in Fintech applications I had previously built AI solutions (using adverse networks) for bias detection within a wider techno-legal framework. But that is a topic for another discussion.
There are other ways to think about the language problem, such as tailoring language to match “brand personality” rather than personality profiles per se. For example, we might want to adapt language that conveys a high trust signal. This type of approach is what start-ups like Persado claim to be pursuing. Of course, it’s possible to combine both approaches: embedding trust signals that are simultaneously oriented towards personality type.
The Long Game
My broader vision for this research was to build a machine that could personalize all touch points across the user journey. This is certainly possible, although we are some way away from fully automated solutions. Meanwhile, AI and language understanding can be used to actively adapt experience. Marketing content to personality type could be implemented in ways not previously possible.
Paul is an expert in applied technology, such as AI and Machine Learning, especially in corporate innovation settings. He is an IEE prize recipient and has over 30 patents, including fields such as computational aesthetics, bias in AI, decentralized finance, natural language generation and computer vision. To read more from Paul check out his website – https://paulgolding.com/