Search engine click logs provide an invaluable source of relevance information, but this information is biased. A key bias presentation order: the probability influenced by a document's position in results page. This paper focuses on explaining that bias, modelling how depends position. We propose four simple hypotheses about might arise. carry out large data-gathering effort, where we perturb ranking major search engine, to see clicks are affected. then explore which best explains...

Reasoning about fairness through correlation-based notions is rife with pitfalls. The 1973 University of California, Berkeley graduate school admissions case from Bickel et. al. (1975) a classic example one such pitfall, namely Simpson's paradox. discrepancy in admission rates among males and female applicants, the aggregate data over all departments, vanishes when per department are examined. We reason causal lens. In process, we introduce statistical test for hypothesis testing based on...

This paper discusses inference problems in probabilistic graphical models that often occur a machine learning setting. In particular it presents unified view of several recently proposed approximation schemes. Expectation consistent approximations and expectation propagation are both shown to be related Bethe free energies with weak consistency constraints, i.e. where local only required agree on certain statistics instead full marginals.

On-street parking, just as any publicly owned utility, is used inefficiently if access free or priced very far from market rates. This paper introduces a novel demand management solution: using data dedicated occupancy sensors an iteration scheme updates parking rates to better match demand. The new encourage parkers avoid peak hours and locations reduce congestion underuse. solution deliberately simple so that it easy understand, easily seen be fair leads policies are remember act upon. We...

We introduce a novel approximate inference algorithm for nonlinear dynamical systems. The is based upon expectation propagation and Gaussian quadrature. first forward pass strongly related to the unscented Kalman filter. It improves filtering by only making approximations in latent not observation space. Smoothed estimates can be found without inverting space dynamics improved iteration. Multiple backward passes make it possible improve local them as consistent possible. demonstrate validity...

We propose a deterministic method to evaluate the integral of positive function based on soft-binning functions that smoothly cut into smaller integrals are easier approximate. In combination with mean-field approximations for each individual sub-part this leads tractable algorithm alternates between optimization bins and approximation local integrals. introduce suitable choices binning such standard mean field can be extended split without need extra derivations. The seen as revival ideas...

We develop a novel multi-armed bandit (MAB) mechanism for the problem of selecting subset crowd workers to achieve an assured accuracy each binary labelling task in cost optimal way. This is challenging because have unknown qualities and strategic costs.

We propose a novel visualization algorithm for high-dimensional time-series data. In contrast to most techniques, we do not assume consecutive data points be independent. The basic model is linear dynamical system which can seen as dynamic extension of probabilistic principal component model. A further particular switching allows representation complex onto multiple and even hierarchy plots. Using sensible approximations based on expectation propagation, the projections performed in...

We describe expectation propagation for approximate inference in dynamic Bayesian networks as a natural extension of Pearl s exact belief propagation.Expectation IS greedy algorithm, converges IN many practical cases, but NOT always.We derive DOUBLE - loop guaranteed TO converge local minimum OF Bethe free energy.Furthermore, we show that stable fixed points (damped) correspond minima this energy, the converse need be CASE .We illustrate algorithms BY applying them switching linear dynamical...

We consider an expert-sourcing problem where the owner of a task benefits from high quality opinions provided by experts. Execution at assured level in cost effective manner becomes mechanism design when individual qualities are private information The considered class execution problems falls into category interdependent values, one cannot simultaneously achieve truthfulness and efficiency unrestricted setting due to impossibility result. propose novel QUEST, that exploits structure our...

Recommender systems have received much attention in recent years, and they been successfully applied many different domains. With each domain come new constraints that require system designers to make choices about how apply extend generic algorithms their context. Booking.com is planet earth's number one accommodation reservation site. The recommendation problem it needs solve has several interesting unique challenges a straightforward matrix factorization or basic bi-linear model are not...

Consider a requester who wishes to crowdsource series of identical binary labeling tasks pool workers so as achieve an assured accuracy for each task, in cost optimal way. The are heterogeneous with unknown but fixed qualities and their costs private. problem is select task subset that the outcome obtained from selected guarantees target level. challenging one even non strategic setting since aggregated label depends on qualities. We develop novel multi-armed bandit (MAB) mechanism solving...