Publications

J. McInerney, N. Kallus. Hessian-Free Laplace in Bayesian Deep Learning. In Workshop on Adaptive Experimental Design and Active Learning in the Real World, Dec 2023. [ pdf ])

N. Kallus.*, J. McInerney*. The Implicit Delta Method. In Neural Information Processing Systems (NeurIPS), Dec 2022. [ pdf ]

The classic delta method has seen a resurgence in deep learning due to its scalability and applicability post-hoc (after training). However, it requires calculating and inverting the Hessian of the objective function which is prohibitive in large models. We develop an implicit delta method that requires just one additional model in the same class as the original fit to a regularized objective.

J. McInerney, N. Kallus. Residual Overfit Method of Exploration. ArXiv, Aug 2021. [ pdf | slides ]

* = equal contribution

J. McInerney*, B. Brost*, P. Chandar*, R. Mehrotra, B. Carterette. Counterfactual Evaluation of Slate Recommendations with Sequential Reward Interactions. In Knowledge Discovery and Data Mining (KDD), Aug 2020. [ pdf ]

A. Gruson, P. Chandar, C. Charbuillet, J. McInerney, S. Hansen, D. Tardieu, B. Carterette. Offline evaluation to make decisions about playlist recommendation. In ACM International Conference on Web Search and Data Mining (WSDM), Feb 2019. [ pdf ]

J. McInerney, B. Lacker, S. Hansen, K. Higley, H. Bouchard, A. Gruson, R. Mehrotra. Explore, exploit, explain: personalizing explainable recommendations with bandits. In Recommender Systems (RecSys), Oct 2018. [ pdf | slides | blog]

Incorporating both exploration and exploitation in recommendation settings (e.g., suggesting movies, songs, products online) in a principled way using contextual bandits and counterfactual training improves performance. This paper proposes a contextual bandit method for explained recommendations, such as would appear on the home page of an online store or streaming platform.

R. Mehrotra, J. McInerney, H. Bouchard, M. Lalmas-Roelleke, F. Diaz. Towards a fair marketplace: counterfactual evaluation of the trade-off between relevance, fairness & satisfaction in recommender systems. In ACM International Conference on Information and Knowledge Management (CIKM) Industry and Case Study Track, Oct 2018 [ pdf ]

D. Liang, L. Charlin, J. McInerney, and D. Blei. Modeling user exposure in recommendation. In International World Wide Web Conference (WWW), April 2016. [ pdf ]

In this paper we explain why downweighting the zeros during training in collaborative filtering, as proposed by Hu et al. 2008, is so effective: it can be interpreted as model with a latent variable for each user-item interaction indicating whether the item was exposed to the user or not. We derive an EM-based algorithm and show how to incorporate side information to exposure.

L. Charlin, R. Ranganath, J. McInerney, and D. Blei. Dynamic Poisson factorization. In Conference on Recommender Systems (RecSys), September 2015. [ pdf ]

Most recommender systems assume static user preferences and item attributes. We propose a model for understanding how users and items evolve over time using Brownian motion in a matrix (Poisson) factorization setting with an efficient algorithm for inference.

* = equal contribution

J. McInerney. An empirical Bayes approach to optimizing machine learning algorithms. In Neural Information Processing Systems (NIPS), Dec 2017. [ pdf ]

Fitting hyperparameters usually requires maximizing the objective w.r.t. hyperparameters on validation data. In this paper, I criticize this method for two reasons: it can overfit hyperparameters and makes the machine learning workflow cumbersome. My alternative, EB-Hyp, avoids these pitfalls using methods inspired by empirical Bayes and Bayesian optimization.

S. Mandt, J. McInerney, F. Abrol, R. Ranganath, and D. Blei. Variational tempering. In International Conference on Artificial Intelligence and Statistics (AISTATS), May 2016. [ pdf ]

The objective in variational inference is non-convex. To avoid local optima, we explore the use of annealing approaches in variational inference. We propose a method called variational tempering to automatically tune the annealing temperature parameter by maximizing likelihood.

J. McInerney, R. Ranganath, and D. Blei. The population posterior and Bayesian modeling on streams. In Neural Information Processing Systems (NIPS), December 2015. [ pdf ]

Most approaches in approximate inference assume a fixed dataset. Streaming data require inference algorithms to constantly update their parameters in an online fashion (also known as never ending learning). Our paper provides a new variational inference approach based on a combination of frequentist and Bayesian concepts. We specify a new variational objective based on a population posterior and show that SVI is a special case where the bootstrap plug-in estimate is used in place of the population distribution.

J. Guerra, J. Uddin, D. Nilsen, J. Mclnerney, A. Fadoo, I. B Omofuma, S. Hughes, S. Agrawal, P. Allen, H. M. Schambra. Capture, learning, and classification of upper extremity movement primitives in healthy controls and stroke patients. In IEEE International Conference on Rehabilitation Robotics, July 2017. [ pdf ]

J. McInerney and D. Blei. Discovering newsworthy tweets with a geographical topic model. In Workshop on Data Science for News Publishing, in conjunction with Conference on Knowledge Discovery and Data Mining (KDD), July 2014. [ pdf ]

Is it possible to automatically discover latent events in a set of documents based on co-occurrence of topics, location, and time? We investigate the possibility using a dataset of tweets with a probabilistic model of events.

J. McInerney, J. Zheng, A. Rogers, and N. R. Jennings. Modelling heterogeneous location habits in human populations for location prediction under data sparsity. In International Joint Conference on Pervasive and Ubiquitous Computing (UbiComp), September 2013. [ pdf ]

Spatio-temporal data are often sparse, yet populations often repeat the same patterns to different degrees. To address the sparsity problem, we propose a mixed membership model of spatio-temporal patterns inspired by topic models and explore the Nokia GPS dataset with it.

A. Rutherford, M. Cebrian, I. Rahwan, S. Dsouza, J. McInerney, V. Naroditskiy, M. Venanzi, N. R. Jennings, J. R. deLara, E. Wahlstedt, and S. U. Miller. Targeted social mobilization in a global manhunt. PLoS ONE, 8(9):e74628, September 2013. [ pdf ]

To what extent can the right incentive structure mobilize people to solve a highly challenging task? The TAG Challenge asked teams to find 5 people hidden in plain sight internationally. Using a hierarchical reward mechanism we were able to win the challenge and we explain how in this paper.

J. McInerney, A. Rogers, and N. R. Jennings. Learning periodic human behaviour models from sparse data for crowdsourcing aid delivery in developing countries. In Conference on Uncertainty in Artificial Intelligence (UAI), pages 401-410, July 2013. [ pdf ]

We consider the problem of combining a probabilistic spatio-temporal model of the world with reinforcement learning. By taking advantage of periodicities in the underlying model, we show how one can make a Markov decision process tractable. The problem of crowdsourcing package delivery is used as a case study of the approach.

J. McInerney, A. Rogers, and N. R. Jennings. Bus, bike and random journeys: crowdsourcing aid distribution in Ivory Coast.Significance, 10(4):4-9, August 2013. [ pdf ]

A write-up of the UAI 2013 paper for a popular audience in the magazine for the Royal Statistical Society.

J. McInerney, S. Stein, A Rogers, and N. R. Jennings. Breaking the habit: measuring and predicting departures from routine in individual human mobility. Pervasive and Mobile Computing, 9(6):808-822, July 2013. [ bib | pdf ]

Anomalies in spatio-temporal data often exist but it is challenging to identify what is an anomaly without knowing what patterns do exist in the data. We consider novel approaches to jointly identifying routines and departures from routines in time series location data. The approach extends the hidden Markov model by adding an additional time series of interacting anomalies.

I. Rahwan, S. Dsouza, A. Rutherford, V. Naroditskiy, J. McInerney, M. Venanzi, N. R. Jennings, and M. Cebrian. Global manhunt pushes the limits of social mobilization. IEEE Computer, 46(4):68-75, April 2013. [ pdf ]

N. Cuong Truong, J. McInerney, L. Tran-Thanh, E. Costanza, and S. D. Ramchurn. Forecasting multi-appliance usage for smart home energy management. In 23rd International Joint Conference on Artificial Intelligence (IJCAI 2013), April 2013. [ pdf ]

J. McInerney, A. Rogers, and N. R. Jennings. Improving location prediction services for new users with probabilistic latent semantic analysis. In 4th International Workshop on Location-Based Social Networks, September 2012. [ pdf ]

J. McInerney, S. Stein, A. Rogers, and N. R. Jennings. Exploring periods of low predictability in daily life mobility. In Mobile Data Challenge by Nokia, July 2012. [ pdf ]