arXiv subjects

Elad Yom-Tov

Publications and source records attributed to Elad Yom-Tov.

28 records · Page 2Linked to original sources

arXiv2017

Evidence from web-based dietary search patterns to the role of B12 deficiency in chronic pain

Profound vitamin B12 deficiency is a known cause of disease, but the role of low or intermediate levels of B12 in the development of neuropathy and other neuropsychiatric symptoms as well as the relationship of eating meat and B12 levels is unclear. Here we use food-related internet search patterns from a sample of 8.5 million US-based people as a proxy to B12 intake and correlate these searches with internet searches related to possible effects of B12 deficiency. Food-related search patterns are highly correlated with known consumption and food-related searches (Spearman 0.69). Awareness of B12 deficiency was associated with a higher consumption of B12-rich foods and with queries for B12 supplements. Searches for terms related to neurological disorders were correlated with searches for B12-poor foods, in contrast with control terms. Popular medicines, those having fewer indications, and those which are predominantly used to treat pain are more strongly correlated with the ability to predict neuropathic pain queries using the B12 contents of food. Our findings provide evidence for the utility of using Internet search patterns to investigate health questions in large populations and suggest that low B12 intake may be associated with a broader spectrum of neurological disorders than currently appreciated.

cs.IR↗

arXiv2017

Modeling influenza-like illnesses through composite compartmental models

Epidemiological models for the spread of pathogens in a population are usually only able to describe a single pathogen. This makes their application unrealistic in cases where multiple pathogens with similar symptoms are spreading concurrently within the same population. Here we describe a method which makes possible the application of multiple single-strain models under minimal conditions. As such, our method provides a bridge between theoretical models of epidemiology and data-driven approaches for modeling of influenza and other similar viruses. Our model extends the Susceptible-Infected-Recovered model to higher dimensions, allowing the modeling of a population infected by multiple viruses. We further provide a method, based on an overcomplete dictionary of feasible realizations of SIR solutions, to blindly partition the time series representing the number of infected people in a population into individual components, each representing the effect of a single pathogen. We demonstrate the applicability of our proposed method on five years of seasonal influenza-like illness (ILI) rates, estimated from Twitter data. We demonstrate that our method describes, on average, 44\% of the variance in the ILI time series. The individual infectious components derived from our model are matched to known viral profiles in the populations, which we demonstrate matches that of independently collected epidemiological data. We further show that the basic reproductive numbers ($R0$) of the matched components are in range known for these pathogens. Our results suggest that the proposed method can be applied to other pathogens and geographies, providing a simple method for estimating the parameters of epidemics in a population.

cs.SI↗

arXiv2017

Automatic Representation for Lifetime Value Recommender Systems

Many modern commercial sites employ recommender systems to propose relevant content to users. While most systems are focused on maximizing the immediate gain (clicks, purchases or ratings), a better notion of success would be the lifetime value (LTV) of the user-system interaction. The LTV approach considers the future implications of the item recommendation, and seeks to maximize the cumulative gain over time. The Reinforcement Learning (RL) framework is the standard formulation for optimizing cumulative successes over time. However, RL is rarely used in practice due to its associated representation, optimization and validation techniques which can be complex. In this paper we propose a new architecture for combining RL with recommendation systems which obviates the need for hand-tuned features, thus automating the state-space representation construction process. We analyze the practical difficulties in this formulation and test our solutions on batch off-line real-world recommendation data.

stat.ML↗

arXiv2016

Predicting drug recalls from Internet search engine queries

Batches of pharmaceutical are sometimes recalled from the market when a safety issue or a defect is detected in specific production runs of a drug. Such problems are usually detected when patients or healthcare providers report abnormalities to medical authorities. Here we test the hypothesis that defective production lots can be detected earlier by monitoring queries to Internet search engines. We extracted queries from the USA to the Bing search engine which mentioned one of 5,195 pharmaceutical drugs during 2015 and all recall notifications issued by the Food and Drug Administration (FDA) during that year. By using attributes that quantify the change in query volume at the state level, we attempted to predict if a recall of a specific drug will be ordered by FDA in a time horizon ranging from one to 40 days in future. Our results show that future drug recalls can indeed be identified with an AUC of 0.791 and a lift at 5% of approximately 6 when predicting a recall will occur one day ahead. This performance degrades as prediction is made for longer periods ahead. The most indicative attributes for prediction are sudden spikes in query volume about a specific medicine in each state. Recalls of prescription drugs and those estimated to be of medium-risk are more likely to be identified using search query data. These findings suggest that aggregated Internet search engine data can be used to facilitate in early warning of faulty batches of medicines.

cs.IR↗

arXiv2016

Predicting Counterfactuals from Large Historical Data and Small Randomized Trials

When a new treatment is considered for use, whether a pharmaceutical drug or a search engine ranking algorithm, a typical question that arises is, will its performance exceed that of the current treatment? The conventional way to answer this counterfactual question is to estimate the effect of the new treatment in comparison to that of the conventional treatment by running a controlled, randomized experiment. While this approach theoretically ensures an unbiased estimator, it suffers from several drawbacks, including the difficulty in finding representative experimental populations as well as the cost of running such trials. Moreover, such trials neglect the huge quantities of available control-condition data which are often completely ignored. In this paper we propose a discriminative framework for estimating the performance of a new treatment given a large dataset of the control condition and data from a small (and possibly unrepresentative) randomized trial comparing new and old treatments. Our objective, which requires minimal assumptions on the treatments, models the relation between the outcomes of the different conditions. This allows us to not only estimate mean effects but also to generate individual predictions for examples outside the randomized sample. We demonstrate the utility of our approach through experiments in three areas: Search engine operation, treatments to diabetes patients, and market value estimation for houses. Our results demonstrate that our approach can reduce the number and size of the currently performed randomized controlled experiments, thus saving significant time, money and effort on the part of practitioners.

cs.LG↗

arXiv2016

Inferring individual attributes from search engine queries and auxiliary information

Internet data has surfaced as a primary source for investigation of different aspects of human behavior. A crucial step in such studies is finding a suitable cohort (i.e., a set of users) that shares a common trait of interest to researchers. However, direct identification of users sharing this trait is often impossible, as the data available to researchers is usually anonymized to preserve user privacy. To facilitate research on specific topics of interest, especially in medicine, we introduce an algorithm for identifying a trait of interest in anonymous users. We illustrate how a small set of labeled examples, together with statistical information about the entire population, can be aggregated to obtain labels on unseen examples. We validate our approach using labeled data from the political domain. We provide two applications of the proposed algorithm to the medical domain. In the first, we demonstrate how to identify users whose search patterns indicate they might be suffering from certain types of cancer. In the second, we detail an algorithm to predict the distribution of diseases given their incidence in a subset of the population at study, making it possible to predict disease spread from partial epidemiological data.

cs.IR↗

arXiv2016

A Reinforcement Learning System to Encourage Physical Activity in Diabetes Patients

Regular physical activity is known to be beneficial to people suffering from diabetes type 2. Nevertheless, most such people are sedentary. Smartphones create new possibilities for helping people to adhere to their physical activity goals, through continuous monitoring and communication, coupled with personalized feedback. We provided 27 sedentary diabetes type 2 patients with a smartphone-based pedometer and a personal plan for physical activity. Patients were sent SMS messages to encourage physical activity between once a day and once per week. Messages were personalized through a Reinforcement Learning (RL) algorithm which optimized messages to improve each participant's compliance with the activity regimen. The RL algorithm was compared to a static policy for sending messages and to weekly reminders. Our results show that participants who received messages generated by the RL algorithm increased the amount of activity and pace of walking, while the control group patients did not. Patients assigned to the RL algorithm group experienced a superior reduction in blood glucose levels (HbA1c) compared to control policies, and longer participation caused greater reductions in blood glucose levels. The learning algorithm improved gradually in predicting which messages would lead participants to exercise. Our results suggest that a mobile phone application coupled with a learning algorithm can improve adherence to exercise in diabetic patients. As a learning algorithm is automated, and delivers personalized messages, it could be used in large populations of diabetic patients to improve health and glycemic control. Our results can be expanded to other areas where computer-led health coaching of humans may have a positive impact.

cs.CY↗

arXiv2015

On the Effect of Human-Computer Interfaces on Language Expression

Language expression is known to be dependent on attributes intrinsic to the author. To date, however, little attention has been devoted to the effect of interfaces used to articulate language on its expression. Here we study a large corpus of text written using different input devices and show that writers unconsciously prefer different letters depending on the interplay between their individual traits (e.g., hand laterality and injuries) and the layout of keyboards. Our results show, for the first time, how the interplay between technology and its users modifies language expression.

cs.HC↗

arXiv2014

Echo chamber amplification and disagreement effects in the political activity of Twitter users

Online social networks have emerged as a significant platform for political discourse. In this paper we investigate what affects the level of participation of users in the political discussion. Specifically, are users more likely to be active when they are surrounded by like-minded individuals, or, alternatively, when their environment is heterogeneous, and so their messages might be carried to people with differing views. To answer this question, we analyzed the activity of about 200K Twitter users who expressed explicit support for one of the candidates of the 2012 US presidential election. We quantified the level of political activity (PA) of users by the fraction of political tweets in their posts, and analyzed the relationship between PA and measures of the users' political environment. These measures were designed to assess the likemindedness, e.g., the fraction of users with similar political views, of their virtual and geographic environments. Our results showed that high PA is usually obtained by users in politically balanced virtual environment. This is in line with the disagreement theory of political science that states that a user's PA is invigorated by the disagreement with their peers. Our results also show that users surrounded by politically like-minded virtual peers tend to have low PA. This observation contradicts the echo chamber amplification theory that states that a person tends to be more politically active when surrounded by like-minded people. Finally, we observe that the likemindedness of the geographical environment does not affect PA. We thus conclude that PA of users is independent of the likemindedness of their geographical environment and is correlated with likemindedness of their virtual environment. The exact form of correlation manifests the phenomenon of disagreement and, in a majority of settings, contradicts the echo chamber amplification theory.

cs.SI↗

arXiv2008

A Gaussian Belief Propagation Solver for Large Scale Support Vector Machines

Support vector machines (SVMs) are an extremely successful type of classification and regression algorithms. Building an SVM entails solving a constrained convex quadratic programming problem, which is quadratic in the number of training samples. We introduce an efficient parallel implementation of an support vector regression solver, based on the Gaussian Belief Propagation algorithm (GaBP). In this paper, we demonstrate that methods from the complex system domain could be utilized for performing efficient distributed computation. We compare the proposed algorithm to previously proposed distributed and single-node SVM solvers. Our comparison shows that the proposed algorithm is just as accurate as these solvers, while being significantly faster, especially for large datasets. We demonstrate scalability of the proposed algorithm to up to 1,024 computing nodes and hundreds of thousands of data points using an IBM Blue Gene supercomputer. As far as we know, our work is the largest parallel implementation of belief propagation ever done, demonstrating the applicability of this algorithm for large scale distributed computing systems.

cs.LG↗