Nth-Iteration

On Friday the 4th of December Maurits gave a talk at the Sensorimotor control lab of Prof. Pieter Medendorp. The slides of the talk on (contextual) Multi Armed Bandit problems can be found Lecture.

The new book Modern Statistical Methods for HCI that Judy Robertson and I (Maurits) are editing is coming up soon; we are now completing all the proofs.

But, to get a sneak preview of some of the content (and just an enjoyable read), check out Judy’s blog: http://judyrobertson.typepad.com/judy_robertson/. Judy is creating a daily post with a nice preview of things that are covered in the book. That should get you to place a preorder 😉

Hi All,

We just moved another version of Streaming Bandit to the master branch; it is quite a bit more mature now, and its properly documented. We are working on the first release. See: https://github.com/MKaptein/streamingbandit.

Hi All,

While working on a project analyzing advertising clicks, we just put together a [R] S4 class to FIT, in a data stream (e.g., row-by-row) a finite mixture of logistic regression models. It can be found under Downloads. Here is an example of it’s use: the code below generates a dataset and plots it. In the final lines we run through the dataset row by row, and fit models with 1, 2, and 3 clusters:

## Usage examples:
source("onlineMixtureLogistic.R")
library(lattice)

# Create a dataset:
set.seed(12345)

n <- 10e4			# Number of subjects
k <- 2				# Number of predictors (including intercept)
j <- 2				# Number of clusters
pj <- c(.3, .7) 	# Cluster probabilties

betas <- matrix( 
			c(
				c(3 , -2.5), 
				c(-2, 5)
			), nrow=j, byrow=TRUE)

X <- matrix(c(rep(1,n),runif((k-1) * n,-5,5)), ncol=k)
cluster <- sample(1:j, n, TRUE, pj)
y <- gen.mixture(X, betas, cluster)

# Plot the dataset:
library(lattice)
xyplot(jitter(y) ~ X[,2], groups=cluster)

# Inspect the elements:
betas
table(cluster) / sum(table(cluster))


# Instantiate object (predictors, clusters)
oLM1 <- OnlineLogMixture(k,1)
oLM2 <- OnlineLogMixture(k,2)
oLM3 <- OnlineLogMixture(k,3)

for(i in 1:nrow(X)){
	oLM1 <- add.observation(oLM1, y[i], X[i,])
	oLM2 <- add.observation(oLM2, y[i], X[i,])
	oLM3 <- add.observation(oLM3, y[i], X[i,])
}

summary(oLM1)
summary(oLM2)
summary(oLM3)

Robin van Emden, who is working in our project on Lock in Feedback (together with Prof. Dr. Davide Iannuzzi from the VU Amsterdam) is blogging. Check out his blog: http://www.pavlov.io.

He will be adding novel results of our Lock in Feedback work – with super cool extensions to the decoy effect – soon. We will keep you updated.

We just added a new downloads section (http://www.nth-iteration.com/downloads/) where we point to resources (software and data) created by the lab.

We are also publishing a dataset of clicks on Persuasive messages. This is a dataset in which the persuasive strategies where selected uniformly randomly, and the dataset can thus be used for the offline evaluation of bandit policies (although, mind the complexities in the data collection – see the README of the download itself).

We hope to add more resources as time goes on.

On the 9th of December Dr. Dean Eckles (MIT Sloan) is visiting!

We are looking forward to his talk entitled: “Estimating Effects in Social Networks with Peer Encouragement Designs”

Here is a brief summary of the talk:

“Interactions among humans enable the spread of information, preferences, behavior, and disease. Despite large-scale measurement of human behaviors, credible identification and estimation of peer influence effects remains difficult. After reviewing other methods for identifying peer effects, including high-dimensional adjustment of observational data and randomized experiments that modulate mechanisms of peer effects, we present research designs that enable point identification of peer effects for relevant populations. In these peer encouragement designs, vertices are randomly assigned to conditions that affect adoption of a target behavior; the experimenter then observes how individuals are affected by the assignment of their peers via their peers’ behaviors. We present an example of a large peer encouragement design that identifies the effects of receiving feedback from peers on Facebook. This experiment reveals “virtuous cycles” in engagement with communication technologies.”

Dean will speak on the 9th of December, at 12.45, in SP.a.007 (Spinoza Building, Radboud University Nijmegen, Nijmegen, the Netherlands). Don’t miss it!

 

Hi,

Recently we have been getting quite a few request for the susceptibility to persuasion scale developed by the lab in 2012. It is cool to see this new spike of interest for this work.

Just for those looking, two references on the use of the scale can be of use:

• http://dl.acm.org/citation.cfm?doid=2209310.2209313
• http://alexandria.tue.nl/extra2/729200.pdf

That should give a full description of the items, the factor solution, the reliability and validity checks, and an example of its use.

Thanks,

Maurits

Ow, and the book on statistical methods that I am co-editing with Judy Robertson just received an ISBN number, which means it exists. You can find some info here.

Hi All,

We are very happy to welcome Robin van Emden to our team. Robin will be working on our joint project with Prof. Davide Iannuzzi from the VU, Amsterdam. We are currently preparing a set of experiments using Lock in Feedback (see here) to determine optimal decoys (see the decoy effect).

Also, I noticed I had not posted these yet: here are the slides of the talk during the Contemporary Challenges in Customized Communication Symposium.

Finally, we are about to launch, I hope in about 2 weeks from now, a first version of StreamingBandit.

Cheers.