For both personality and preference vectors, it would be great to see data from the data from OKCupid, from its good old days (<a href="https://www.reddit.com/r/gwern/comments/aapn1l/okcupid_blog_archives/" rel="nofollow">https://www.reddit.com/r/gwern/comments/aapn1l/okcupid_blog_...</a>).<p>Even more, since for questions there are:<p>- questions one decided to answer (which says something on its own)<p>- answers<p>- declared answers they accept in their partners<p>- actual partners they pursue (judged by matches, or dates)<p>While I expect that mostly similarities attract (so-called associative mating), there are compatible traits (e.g one person loves to listen, one person loves to talk), and there is the level of lack of self-knowledge, or hypocrisy (what we SAY we like, vs what we actually do).<p>And then e.g. probability that person A likes person B can be expressed as:<p>sigmoid(actualPrefVecA * personalityVecB)<p>...and with gradient descent magic, we can turn people into vectors!
It's fun to see folks rederive these things. Multiple arms of psychology have dove deep into this for many years now, including social psychology (individual biases and relationship behaviors), cognitive (modeling decision making processes), quantitative psychometrics (formal mathematical models of how people represent abstract concepts or traits), personality psychology (emphasis on individual differences and patterns), and of course, clinical psych.<p>Lots of overlap among these, but rather than start from scratch, a bit of reading in almost any of them, perhaps starting with behavioral economics or social psych, might enhance the "vectors".
I was recently thinking about this, but on a slightly bigger scale.<p>There’s a term “off-kilter”, which is easy to explain using vectors like this.<p>If we take the general vector of society, just sum up all personal vectors and normalise, we form a big vector for society. This is what’s “normal”.<p>Kilter refers to the concept of how aligned we are with society’s vector. So the concept of being off-kilter is how skewed you are wrt this normal vector essentially.<p>Of course valid for any of the eigenvectors corresponding to subfields again, and this also goes some way to help form the overton window, which has recently been up here in some posts...<p>It’s a fun sociomathematical formulation.
> You can take these preferences and combine them into a single value/point on a vector.<p>No, you can't.<p>Also, people don't have "preferences" in the way the article posits
It'd be nice if vector based approaches (think MBTI tests being originally created using PCA) would use state of the art dimensionality reduction techniques instead of stuff from the 70s so that way each vector actually matters and explains far more variance in the data.<p>If you need interpretability just use an autoencoder or simese network instead of PCA. This way the new "personality" vectors are highly meaningful and visualizations are better...
I have thought of Personality Vectors similar to this concept but for a completely different purpose.<p>If we had a Chatbot that would interact with someone hypothetically we could have this kind of distinction between Personality Vectors to mutate its own behavior and to interact with others.<p>Its sort of the idea I had when I watched the movie Interstellar and TARS is told to decrease his own humor. So if the system would just know who its talking to it would create a preference vector for each person it interacts with.<p>But you would need some kind of baseline so basically the Personality Vectors would be set to neutral. And each interaction with a Personality vector would be a system that stores a preference vector for each person.<p>As the system time evolves (which is also in the article) you would basically end up with a set of graphs. There is a notion out there which is called a Dynamic Network analysis [1]. If you also stored the history of that you could then not just be able to rewind the system backwards to a previous checkpoint.<p>The way you could start learning how to learn a personality vector is a reinforcement learning situation. If you say "bad AI I didn't like that" the system would then look at what it said and mutate the preference vector for that person. Or you could just change the vector manually by a technique that was illustrated above by just saying "reduce humor by 70%".<p>Then you could have model which would do conditional generation [1] of text based upon interactions with others. It would be basically a reinforcement learning system paired with a generative text model. The reinforcement system would store the Personality and Preference vectors for each person. For each behavior and preference you would need a corpus to bootstrap the system. The reinforcement system would be a retrieval network and it would retrieve a pertained system based on user(s) input and get a generational text network to provide a response. This would be similar to Alpha Go's design. [3]<p>[1] <a href="https://en.wikipedia.org/wiki/Dynamic_network_analysis" rel="nofollow">https://en.wikipedia.org/wiki/Dynamic_network_analysis</a>
[2] <a href="https://github.com/salesforce/ctrl" rel="nofollow">https://github.com/salesforce/ctrl</a>
[3] <a href="https://datascience.stackexchange.com/questions/10932/difference-between-alphagos-policy-network-and-value-network" rel="nofollow">https://datascience.stackexchange.com/questions/10932/differ...</a>
Interesting article, and good timing - A few days ago I wrote about how we can build decentralized social networks where you can filter your world to see content from those with similar personality vectors to yourself - <a href="https://adecentralizedworld.com/2020/06/a-trust-and-moderation-system-for-the-decentralized-web/" rel="nofollow">https://adecentralizedworld.com/2020/06/a-trust-and-moderati...</a>
The concept of ordered preferences is the foundation of Microeconomics and Game theory.<p>For those of you who are interested, a good introduction can be found here : <a href="https://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.307.4118&rep=rep1&type=pdf" rel="nofollow">https://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.30...</a>
The idea of preferences as vectors seems like it has some fun consequences.<p>- HIPSTERS: If you really like something, but don't want others to crowd your interests, you will pretend to be nonchalant about it.<p>- AUTHENTICITY: A person's interests in a topic can be vetted as "authentic" if they are influenced by others whose average of personality vectors matches the person's.<p>- SOCIOPATHS: You can make in-roads with a group of interest by transforming your PV; see AUTHENTICITY.<p>- SKEW: You can influence others more by projecting a _very_ high value for a specific topic. If you love a particular form of music to an absurd degree, you're more likely to convert others.<p>I don't have any non-anecdotal evidence of any of these phenomena. Rather, these are some reasonable if not amusing deductions from the author's model.