Preference Elicitation on Motor Trends Dataset

This vignette demonstrates how to use the prefeR package on a real dataset. The mtcars dataset provides us such an opportunity.

1974 Motor Trends Car Data
	mpg	cyl	disp	hp	drat	wt	qsec	vs	am	gear	carb
Mazda RX4	21.0	6	160	110	3.90	2.620	16.46	0	1	4	4
Mazda RX4 Wag	21.0	6	160	110	3.90	2.875	17.02	0	1	4	4
Datsun 710	22.8	4	108	93	3.85	2.320	18.61	1	1	4	1
Hornet 4 Drive	21.4	6	258	110	3.08	3.215	19.44	1	0	3	1
Hornet Sportabout	18.7	8	360	175	3.15	3.440	17.02	0	0	3	2

If we wanted to give a user a list of their top five most preferred cars from the mtcars dataset, there are three approaches we could take:

Have our user manually rank all options.
Make the user provide weights for the desirability of different car features, and calculate the weighted value of each option.
Have the user compare a small number of alternatives, and derive their weights from those comparisons.

Option #1 quickly becomes an enormous burden on the user as the number of alternatives increases. Option #2 is difficult for the user to do and replicate. What exactly does it mean if the weight assigned to horsepower is double the weight assigned to fuel efficiency?

Option #3 is enabled by the preference elicitation package. To begin, we create a preference elicitation object and give it our data:

library(prefeR)
p <- prefEl(data = mtcars)
p
## Preference elicitation object with:
##  32 observations of 11 variables.
## And the following preferences:
##  No strict preferences.
##  No indifference preferences.

Now we can add in our Bayesian priors for the weights. Although it is difficult to determine weights exactly, usually one has some ballpark estimate for what they should be, and often one knows with certainty the sign of the weights: all else equal, everyone would prefer a more fuel efficient car. The prefeR package contains three built-in priors:

Normal(mu, sigma) provides a one-dimensional Normal prior with mean mu and standard deviation sigma. This prior is useful if you have a good guess for what the weight should be, and an understanding of how much you expect to differ from that guess.
Exp(mu) provides a one dimensional Exponential prior with mean mu (not rate!). This prior is particularly useful if you deterministically know the sign of the weight, and have a guess for the value of the weight. The mean may be negative.
Flat() yields a completely agnostic, flat prior.

We can now add in our priors for our mtcars attributes.

p$priors <- c(Exp(1),   # MPG
              Normal(), # Number of cylinders (Normal() = Normal(0, 1))
              Normal(), # displacement
              Exp(2),   # horsepower
              Normal(), # real axle ratio
              Normal(), # weight
              Exp(-3),  # quarter mile time
              Normal(), # Engine type
              Normal(), # transmission type
              Normal(), # number of gears
              Normal()  # number of carburetors
)

Now, we can add in our user’s preferences:

p$addPref("Pontiac Firebird" %>% "Fiat 128")  # prefer a cool sports car 
p$addPref("Mazda RX4 Wag" %<% "Mazda RX4")    # prefer not to have the station wagon
p$addPref("Merc 280" %=% "Merc 280C")         # indifferent about C-option

p
## Preference elicitation object with:
##  32 observations of 11 variables.
## And the following preferences:
##  Pontiac Firebird preferred to Fiat 128
##  Mazda RX4 preferred to Mazda RX4 Wag
##  Merc 280 indifferent to Merc 280C

Now, we can infer what our attribute weights should be:

p$infer()
##          mpg          cyl         disp           hp         drat           wt 
##  0.000000000  0.005892875  0.473346403  0.000000000 -0.001473219 -0.286931232 
##         qsec           vs           am         gear         carb 
## -0.284793197 -0.001473219 -0.001473219 -0.001473219  0.001473219

And we can get our top five cars:

p$rank()[1:5]
##  Cadillac Fleetwood Lincoln Continental   Chrysler Imperial    Pontiac Firebird 
##            216.8368            211.1522            201.8215            183.4207 
##          Duster 360 
##            164.9131

Finally, we can figure out what query we should answer next:

p$suggest()
## [1] "Pontiac Firebird"   "Cadillac Fleetwood"

John Lepird

2022-04-24