EXPRESS: Bayesian Consumer Profiling: How to Estimate Consumer Characteristics from Aggregate Data

Firms use aggregate data from data brokers (e.g., Acxiom, Experian) and external data sources (e.g., Census) to infer the likely characteristics of consumers in a target list and thus better predict consumers’ profiles and needs unobtrusively. We demonstrate that the simple count method most commonly used in this effort relies implicitly on an assumption of conditional independence that fails to hold in many settings of managerial interest. We develop a Bayesian profiling introducing different conditional independence assumptions. We also show how to introduce additional observed covariates into this model. We use simulations to show that in managerially relevant settings, the Bayesian method will outperform the simple count method, often by an order of magnitude. We then compare different conditional independence assumptions in two case studies. The first example estimates customers’ age on the basis of their first names; prediction errors decrease substantially. In the second example, we infer the income, occupation, and education of online visitors of a marketing analytic software company based exclusively on their IP addresses. The face validity of the predictions improves dramatically and reveals an interesting (and more complex) endogenous list-selection mechanism than the one suggested by the simple count method.

When alternative hypotheses shape your beliefs: Context effects in probability judgments

When people are asked to estimate the probability of an event occurring, they sometimes make different subjective probability (SP) judgments for different descriptions of the same event. This implies the evidence or support recruited to make SPs is based on the descriptions of the events (hypotheses) instead of the events themselves, as captured by Tversky and Koehler's (1994) support theory. However, is the support assigned to a hypothesis invariant, as support theory assumes? Here, across two studies where participants were asked to estimate the probability that an event would occur, we show that the support people recruit about the target hypothesis also depends on the other hypotheses under consideration. The first study shows that the presence of a distractor---a hypothesis objectively dominated by the target hypothesis---boosts the SP assigned to the target hypothesis. The second study shows that the presence of a resembler---a hypothesis that is objectively similar to the target hypothesis---detracts more from the SP assigned to the target hypothesis than the competing hypothesis. These context effects invalidate the regularity and the strong independence assumptions of support theory and more generally suggest a similar process that drives the construction of preference also underlies belief.

Recovering Secrets From Prefix-Dependent Leakage

We discuss how to recover a secret bitstring given partial information obtained during a computation over that string, assuming the computation is a deterministic algorithm processing the secret bits sequentially. That abstract situation models certain types of side-channel attacks against discrete logarithm and RSA-based cryptosystems, where the adversary obtains information not on the secret exponent directly, but instead on the group or ring element that varies at each step of the exponentiation algorithm.Our main result shows that for a leakage of a single bit per iteration, under suitable statistical independence assumptions, one can recover the whole secret bitstring in polynomial time. We also discuss how to cope with imperfect leakage, extend the model to k-bit leaks, and show how our algorithm yields attacks on popular cryptosystems such as (EC)DSA.

Discrete-State Variational Autoencoders for Joint Discovery and Factorization of Relations

We present a method for unsupervised open-domain relation discovery. In contrast to previous (mostly generative and agglomerative clustering) approaches, our model relies on rich contextual features and makes minimal independence assumptions. The model is composed of two parts: a feature-rich relation extractor, which predicts a semantic relation between two entities, and a factorization model, which reconstructs arguments (i.e., the entities) relying on the predicted relation. The two components are estimated jointly so as to minimize errors in recovering arguments. We study factorization models inspired by previous work in relation factorization and selectional preference modeling. Our models substantially outperform the generative and agglomerative-clustering counterparts and achieve state-of-the-art performance.