Data Skeptic

DataRefuge is a public collaborative, grassroots effort around the United States in which scientists, researchers, computer scientists, librarians and other volunteers are working to download, save, and re-upload government data. The DataRefuge Project, which is led by the UPenn Program in Environmental Humanities and the Penn Libraries group at University of Pennsylvania, aims to foster resilience in an era of anthropogenic global climate change and raise awareness of how social and political events affect transparency.

If a CEO wants to know the state of their business, they ask their highest ranking executives. These executives, in turn, should know the state of the business through reports from their subordinates. This structure is roughly analogous to a process observed in deep learning, where each layer of the business reports up different types of observations, KPIs, and reports to be interpreted by the next layer of the business. In deep learning, this process can be thought of as automated feature engineering. DNNs built to recognize objects in images may learn structures that behave like edge detectors in the first hidden layer. Proceeding layers learn to compose more abstract features from lower level outputs. This episode explore that analogy in the context of automated feature engineering. Linh Da and Kyle discuss a particular image in this episode. The image included below in the show notes is drawn from the work of Lee, Grosse, Ranganath, and Ng in their paper Convolutional Deep Belief Networks for Scalable Unsupervised Learning of Hierarchical Representations.

In this episode, I speak with Raghu Ramakrishnan, CTO for Data at Microsoft. We discuss services, tools, and developments in the big data sphere as well as the underlying needs that drove these innovations.

In this episode, we talk about a high-level description of deep learning. Kyle presents a simple game (pictured below), which is more of a puzzle really, to try and give Linh Da the basic concept. Thanks to our sponsor for this week, the Data Science Association. Please check out their upcoming Dallas conference at dallasdatascience.eventbrite.com

Versioning isn't just for source code. Being able to track changes to data is critical for answering questions about data provenance, quality, and reproducibility. Daniel Whitenack joins me this week to talk about these concepts and share his work on Pachyderm. Pachyderm is an open source containerized data lake. During the show, Daniel mentioned the Gopher Data Science github repo as a great resource for any data scientists interested in the Go language. Although we didn't mention it, Daniel also did an interesting analysis on the 2016 world chess championship that complements our recent episode on chess well. You can find that post here Supplemental music is Lee Rosevere's Let's Start at the Beginning. Thanks to Periscope Data for sponsoring this episode. More about them at periscopedata.com/skeptics

Logistic Regression is a popular classification algorithm. In this episode, we discuss how it can be used to determine if an audio clip represents one of two given speakers. It assumes an output variable (isLinhda) is a linear combination of available features, which are spectral bands in the discussion on this episode. Keep an eye on the dataskeptic.com blog this week as we post more details about this project. Thanks to our sponsor this week, the Data Science Association. Please check out their upcoming conference in Dallas on Saturday, February 18th, 2017 via the link below. dallasdatascience.eventbrite.com

Prior work has shown that people's response to competition is in part predicted by their gender. Understanding why and when this occurs is important in areas such as labor market outcomes. A well structured study is challenging due to numerous confounding factors. Peter Backus and his colleagues have identified competitive chess as an ideal arena to study the topic. Find out why and what conclusions they reached. Our discussion centers around Gender, Competition and Performance: Evidence from Real Tournaments from Backus, Cubel, Guid, Sanchez-Pages, and Mañas. A summary of their paper can also be found here.

Deep learning can be prone to overfit a given problem. This is especially frustrating given how much time and computational resources are often required to converge. One technique for fighting overfitting is to use dropout. Dropout is the method of randomly selecting some neurons in one's network to set to zero during iterations of learning. The core idea is that each particular input in a given layer is not always available and therefore not a signal that can be relied on too heavily.

In this episode I speak with Clarence Wardell and Kelly Jin about their mutual service as part of the White House's Police Data Initiative and Data Driven Justice Initiative respectively. The Police Data Initiative was organized to use open data to increase transparency and community trust as well as to help police agencies use data for internal accountability. The PDI emerged from recommendations made by the Task Force on 21st Century Policing. The Data Driven Justice Initiative was organized to help city, county, and state governments use data-driven strategies to help low-level offenders with mental illness get directed to the right services rather than into the criminal justice system.

We close out 2016 with a discussion of a basic interview question which might get asked when applying for a data science job. Specifically, how a library might build a model to predict if a book will be returned late or not.

Today's episode is a reading of Isaac Asimov's Franchise. As mentioned on the show, this is just a work of fiction to be enjoyed and not in any way some obfuscated political statement. Enjoy, and happy holidays!

Classically, entropy is a measure of disorder in a system. From a statistical perspective, it is more useful to say it's a measure of the unpredictability of the system. In this episode we discuss how information reduces the entropy in deciding whether or not Yoshi the parrot will like a new chew toy. A few other everyday examples help us examine why entropy is a nice metric for constructing a decision tree.

Cloud services are now ubiquitous in data science and more broadly in technology as well. This week, I speak to Mark Souza, Tobias Ternström, and Corey Sanders about various aspects of data at scale. We discuss the embedding of R into SQLServer, SQLServer on linux, open source, and a few other cloud topics.

Today's episode is all about Causal Impact, a technique for estimating the impact of a particular event on a time series. We talk to William Martin about his research into the impact releases have on app and we also chat with Karen Blakemore about a project she helped us build to explore the impact of a Saturday Night Live appearance on a musician's career. Martin's work culminated in a paper Causal Impact for App Store Analysis. A shorter summary version can be found here. His company helping app developers do this sort of analysis can be found at crestweb.cs.ucl.ac.uk/appredict/.

The Bootstrap is a method of resampling a dataset to possibly refine it's accuracy and produce useful metrics on the result. The bootstrap is a useful statistical technique and is leveraged in Bagging (bootstrap aggregation) algorithms such as Random Forest. We discuss this technique related to polling and surveys.

The Gini Coefficient (as it relates to decision trees) is one approach to determining the optimal decision to introduce which splits your dataset as part of a decision tree. To pick the right feature to split on, it considers the frequency of the values of that feature and how well the values correlate with specific outcomes that you are trying to predict.

Financial analysis techniques for studying numeric, well structured data are very mature. While using unstructured data in finance is not necessarily a new idea, the area is still very greenfield. On this episode,Delia Rusu shares her thoughts on the potential of unstructured data and discusses her work analyzing Wikipedia to help inform financial decisions. Delia's talk at PyData Berlin can be watched on Youtube (Estimating stock price correlations using Wikipedia). The slides can be found here and all related code is available on github.

AdaBoost is a canonical example of the class of AnyBoost algorithms that create ensembles of weak learners. We discuss how a complex problem like predicting restaurant failure (which is surely caused by different problems in different situations) might benefit from this technique.

Platform as a service is a growing trend in data science where services like fraud analysis and face detection can be provided via APIs. Such services turn the actual model into a black box to the consumer. But can the model be reverse engineered? Florian Tramèr shares his work in this episode showing that it can. The paper Stealing Machine Learning Models via Prediction APIs is definitely worth your time to read if you enjoy this episode. Related source code can be found in https://github.com/ftramer/Steal-ML.

For machine learning models created with the random forest algorithm, there is no obvious diagnostic to inform you which features are more important in the output of the model. Some straightforward but useful techniques exist revolving around removing a feature and measuring the decrease in accuracy or Gini values in the leaves. We broadly discuss these techniques in this episode.

As cities provide bike sharing services, they must also plan for how to redistribute bicycles as they inevitably build up at more popular destination stations. In this episode, Hui Xiong talks about the solution he and his colleagues developed to rebalance bike sharing systems.

Random forest is a popular ensemble learning algorithm which leverages bagging both for sampling and feature selection. In this episode we make an analogy to the process of running a bookstore.

Jo Hardin joins us this week to discuss the ASA's Election Prediction Contest. This is a competition aimed at forecasting the results of the upcoming US presidential election competition. More details are available in Jo's blog post found here. You can find some useful R code for getting started automatically gathering data from 538 via Jo's github and official contest details are available here. During the interview we also mention Daily Kos and 538.

The F1 score is a model diagnostic that combines precision and recall to provide a singular evaluation for model comparison. In this episode we discuss how it applies to selecting an interior designer.

Urban congestion effects every person living in a city of any reasonable size. Lewis Lehe joins us in this episode to share his work on downtown congestion pricing. We explore topics of how different pricing mechanisms effect congestion as well as how data visualization can inform choices. You can find examples of Lewis's work at setosa.io. His paper which we discussed during the interview isDistance-dependent congestion pricing for downtown zones. On this episode, we discuss State of California data which can be found at pems.dot.ca.gov.

Heteroskedasticity is a term used to describe a relationship between two variables which has unequal variance over the range. For example, the variance in the length of a cat's tail almost certainly changes (grows) with age. On the other hand, the average amount of chewing gum a person consume probably has a consistent variance over a wide range of human heights. We also discuss some issues with the visualization shown in the tweet embedded below.

Our guest today is Michael Cuthbert, an associate professor of music at MIT and principal investigator of the Music21 project, which we focus our discussion on today. Music21 is a python library making analysis of music accessible and fun. It supports integration with popular formats such as MIDI, MusicXML, Lilypond, and others. It's also well integrated with The Elvis Project, enabling users to import large volumes of music for easy analysis. Music21 is a great platform for musicologists and machine learning researchers alike to explore patterns and structure in music.

Paxos is a protocol for arriving a consensus in a distributed computing system which accounts for unreliability of the nodes. We discuss how this might be used in the real world in the event of a massive disaster.

Machine learning models are often criticized for being black boxes. If a human cannot determine why the model arrives at the decision it made, there's good cause for skepticism. Classic inspection approaches to model interpretability are only useful for simple models, which are likely to only cover simple problems. The LIME project seeks to help us trust machine learning models. At a high level, it takes advantage of local fidelity. For a given example, a separate model trained on neighbors of the example are likely to reveal the relevant features in the local input space to reveal details about why the model arrives at it's conclusion. In this episode, Marco Tulio Ribeiro joins us to discuss how LIME (Locally Interpretable Model-Agnostic Explanations) can help users trust machine learning models. The accompanying paper is titled "Why Should I Trust You?": Explaining the Predictions of Any Classifier.

Analysis of variance is a method used to evaluate differences between the two or more groups. It works by breaking down the total variance of the system into the between group variance and within group variance. We discuss this method in the context of wait times getting coffee at Starbucks.

When humans describe images, they have a reporting bias, in that the report only what they consider important. Thus, in addition to considering whether something is present in an image, one should consider whether it is also relevant to the image before labeling it. Ishan Misra joins us this week to discuss his recent paper Seeing through the Human Reporting Bias: Visual Classifiers from Noisy Human-Centric Labels which explores a novel architecture for learning to distinguish presence and relevance. This work enables web-scale datasets to be useful for training, not just well groomed hand labeled corpora.

Survival analysis techniques are useful for studying the longevity of groups of elements or individuals, taking into account time considerations and right censorship. This episode explores how survival analysis can describe marriages, in particular, using the non-parametric Cox proportional hazard model. This episode discusses some good summaries of survey data on marriage and divorce which can be found here. The python lifelines library is a good place to get started for people that want to do some hands on work.

This week is an insightful discussion with Claudia Perlich about some situations in machine learning where models can be built, perhaps by well-intentioned practitioners, to appear to be highly predictive despite being trained on random data. Our discussion covers some novel observations about ROC and AUC, as well as an informative discussion of leakage. Much of our discussion is inspired by two excellent papers Claudia authored: Leakage in Data Mining: Formulation, Detection, and Avoidance and On Cross Validation and Stacking: Building Seemingly Predictive Models on Random Data. Both are highly recommended reading!

An ROC curve is a plot that compares the trade off of true positives and false positives of a binary classifier under different thresholds. The area under the curve (AUC) is useful in determining how discriminating a model is. Together, ROC and AUC are very useful diagnostics for understanding the power of one's model and how to tune it.

I'm joined by Chris Stucchio this week to discuss how deliberate or uninformed statistical practitioners can derive spurious and arbitrary results via multiple comparisons. We discuss p-hacking and a variety of other important lessons and tips for proper analysis. You can enjoy Chris's writing on his blog at chrisstucchio.com and you may also like his recent talk Multiple Comparisons: Make Your Boss Happy with False Positives, Guarenteed.

If you'd like to make a good prediction, your best bet is to invent a time machine, visit the future, observe the value, and return to the past. For those without access to time travel technology, we need to avoid including information about the future in our training data when building machine learning models. Similarly, if any other feature whose value would not actually be available in practice at the time you'd want to use the model to make a prediction, is a feature that can introduce leakage to your model.

Kristian Lum (@KLdivergence) joins me this week to discuss her work at @hrdag on predictive policing. We also discuss Multiple Systems Estimation, a technique for inferring statistical information about a population from separate sources of observation. If you enjoy this discussion, check out the panel Tyranny of the Algorithm? Predictive Analytics & Human Rights which was mentioned in the episode.

Distributed computing cannot guarantee consistency, accuracy, and partition tolerance. Most system architects need to think carefully about how they should appropriately balance the needs of their application across these competing objectives. Linh Da and Kyle discuss the CAP Theorem using the analogy of a phone tree for alerting people about a school snow day.

A startup is claiming that they can detect terrorists purely through facial recognition. In this solo episode, Kyle explores the plausibility of these claims.

Goodhart's law states that "When a measure becomes a target, it ceases to be a good measure". In this mini-episode we discuss how this affects SEO, call centers, and Scrum.

I'm joined this week by Jon Morra, director of data science at eHarmony to discuss a variety of ways in which machine learning and data science are being applied to help connect people for successful long term relationships. Interesting open source projects mentioned in the interview include Face-parts, a web service for detecting faces and extracting a robust set of fiducial markers (features) from the image, and Aloha, a Scala based machine learning library. You can learn more about these and other interesting projects at the eHarmony github page. In the wrap up, Jon mentioned the LA Machine Learning meetup which he runs. This is a great resource for LA residents separate and complementary to datascience.la groups, so consider signing up for all of the above and I hope to see you there in the future.

Mystery shoppers and fruit cultivation help us discuss stationarity - a property of some time serieses that are invariant to time in several ways. Differencing is one approach that can often convert a non-stationary process into a stationary one. If you have a stationary process, you get the benefits of many known statistical properties that can enable you to do a significant amount of inferencing and prediction.

I'm joined by Wes McKinney (@wesmckinn) and Hadley Wickham (@hadleywickham) on this episode to discuss their joint project Feather. Feather is a file format for storing data frames along with some metadata, to help with interoperability between languages. At the time of recording, libraries are available for R and Python, making it easy for data scientists working in these languages to quickly and effectively share datasets and collaborate.

Bargaining is the process of two (or more) parties attempting to agree on the price for a transaction. Game theoretic approaches attempt to find two strategies from which neither party is motivated to deviate. These strategies are said to be in equilibrium with one another. The equilibriums available in bargaining depend on the the transaction mechanism and the information of the parties. Discounting (how long parties are willing to wait) has a significant effect in this process. This episode discusses some of the choices Kyle and Linh Da made in deciding what offer to make on a house.

Deepjazz is a project from Ji-Sung Kim, a computer science student at Princeton University. It is built using Theano, Keras, music21, and Evan Chow's project jazzml. Deepjazz is a computational music project that creates original jazz compositions using recurrent neural networks trained on Pat Metheny's "And Then I Knew". You can hear some of deepjazz's original compositions on soundcloud.

When working with time series data, there are a number of important diagnostics one should consider to help understand more about the data. The auto-correlative function, plotted as a correlogram, helps explain how a given observations relates to recent preceding observations. A very random process (like lottery numbers) would show very low values, while temperature (our topic in this episode) does correlate highly with recent days. See the show notes with details about Chapel Hill, NC weather data by visiting: https://dataskeptic.com/blog/episodes/2016/acf-correlograms

This week I spoke with Elham Shaabani and Paulo Shakarian (@PauloShakASU) about their recent paper Early Identification of Violent Criminal Gang Members (also available onarXiv). In this paper, they use social network analysis techniques and machine learning to provide early detection of known criminal offenders who are in a high risk group for committing violent crimes in the future. Their techniques outperform existing techniques used by the police. Elham and Paulo are part of the Cyber-Socio Intelligent Systems (CySIS) Lab.

A dinner party at Data Skeptic HQ helps teach the uses of fractional factorial design for studying 2-way interactions.

Cheng-tao Chu (@chengtao_chu) joins us this week to discuss his perspective on common mistakes and pitfalls that are made when doing machine learning. This episode is filled with sage advice for beginners and intermediate users of machine learning, and possibly some good reminders for experts as well. Our discussion parallels his recent blog postMachine Learning Done Wrong. Cheng-tao Chu is an entrepreneur who has worked at many well known silicon valley companies. His paper Map-Reduce for Machine Learning on Multicore is the basis for Apache Mahout. His most recent endeavor has just emerged from steath, so please check out OneInterview.io.

Co-host Linh Da was in a biking accident after hitting a pothole. She sustained an injury that required stitches. This is the story of our quest to file a 311 complaint and track it through the City of Los Angeles's open data portal. My guests this episode are Chelsea Ursaner (LA City Open Data Team), Ben Berkowitz (CEO and founder of SeeClickFix), and Russ Klettke (Editor of pothole.info)