PyData London 2019
2019-07-12..2019-07-14
London, United Kingdom
[PyData London 2016, PyData London 2017, PyData London 2018, PyData London 2019, PyData London 2022, PyData London 2024, PyData London 2025, PyData London 2026]
| title | author | length |
|---|---|---|
| The Turing Way: A how to guide for reproducible research | Kirstie Whitaker | 33m |
The Turing Way claims a practical, community‑authored handbook that reduces reproducible research to concrete steps: version control workflows, automated tests… | ||
| Bridging health inequalities through machine learning | Elina Naydenova | 39m |
Describes Feebris’ ML-powered health platform that pairs digital stethoscope + pulse oximeter with a mobile app, tested in a 1300‑child 2017 India trial and… | ||
| Goofing Off with Fun (Big) Data | Lynn Cherny | 51m |
Builds two playable demos and reports what actually works: a Bosch-bot heatmap of most-liked zooms (data pulled with twint + pandas scrape, image registration… | ||
| The frustration of diversity efforts in STEM | Lorena A. Barba | 53m |
Argues that claimed biological causes for low female representation lack supporting leaps and that common fixes fail: implicit-bias training shows little… | ||
| Build and Deploy an End-to-end Streaming NLP Insight System | Michal Mucha | 01h 25m |
Builds a working end to end pipeline that ingests live Reddit comments, serializes them to JSON, pushes them through uses Redis as message queue, runs spaCy… | ||
| Maintainable code in data science | Kevin Lemagnen | 01h 28m |
Notebooks refactor into a testable Python package that builds a full pipeline: custom preprocessing classes that memorize state, a ColumnTransformer + Pipeline… | ||
| Data Analysis in Parallel | Filip Ter | 56m |
Shows how to take a Jupyter workflow and run three real tasks — aggregate stats, a top-20 lemmas histogram, and a sentiment pass using a serialized Keras h5… | ||
| Adv. Software Testing for Data Scientists | Raoul-Gabriel Urma, Kevin Lemagnen | 01h 29m |
Workshop demonstrates concrete practices: use pytest for low-boilerplate unit tests, numpy/pandas-specific assertions like pandas.assert_frame_equal… | ||
| Mastering gradient boosting with CatBoost | Anna Veronika Dorogush | 01h 28m |
CatBoost demonstrates concrete wins: out‑of‑the‑box categorical feature support, defaults that give strong quality without heavy tuning, and benchmarks where… | ||
| Choose the right neural generative model for your problem | Egor Kraev | 44m |
Advice hinges on concrete tradeoffs: pick models by output structure and whether an objective exists. Continuous outputs favor direct nets or autoencoders for… | ||
| An introduction to Markov Chain Monte Carlo using PyMC3 | Chris Fonnesbeck | 01h 30m |
Covers concrete MCMC practice using PyMC3 and the repo https://github.com/fonnesbeck/mcmc_pydata_london_2019 with runnable notebooks: hand-implements… | ||
| AB-testing by cluster | Bertil Hatt, João Martins | 01h 15m |
Concrete claims and code: Python notebooks that run t-test based A/B and A/A batch simulations, compute p-values and confidence intervals, and measure minimum… | ||
| A gentle introduction to Pandas timeseries and Seaborn | Ian Ozsvald | 01h 24m |
Uses Pandas and Seaborn in a Jupyter/Anaconda stack to manipulate a German utilities CSV (~4,383 rows over ~2006–2017) and shows concrete code: to_datetime… | ||
| The anatomy of a #deepfake | Eric Drass | 42m |
A first-hand exploration of the world of #deepfake and post-truth technologies, looking at the latest developments in art and ML-modelling. Including details… | ||
| Guide towards algorithm explainability in machine learning | Alejandro Saucedo | 39m |
Practical walkthrough that trains a loan-approval model on ~8,000 rows, gets 98% vs 54% when tested on production-like labels, and then applies concrete steps… | ||
| Data Science for Dinner: Recommending Recipes | Irene Iriarte Carretero | 35m |
Gousto implements a pragmatic hybrid recommender: a failed attempt with LightFM matrix factorization leads to a content-based system built on a recipe taxonomy… | ||
| 3 things about Chatbot (you don't have to learn it the hard way) | Cheuk Ting Ho | 38m |
Three practical takeaways: a chatbot needs NLP, a dialog manager, scripted content and integrations and those are the things that take most time to build not… | ||
| Active learning in the interactive python environment | Jan Freyberg | 39m |
Claims active learning cuts labels roughly in half — e.g. reduced labels from 10,000 to 4,000 and 7,600 to 3,600 on image tasks. Implementation uses Jupyter +… | ||
| Embeddings! Embeddings everywhere! | Maciej Arciuch, Karol Grzegorczyk | 38m |
Allegro describes a production recommender for ~100M offers and 20M accounts that learns item vectors with prod2vec using skip-gram, tunes hyperparameters to… | ||
| How to Validate you Client Churn Model | Elena Sharova | 31m |
Claims a survivals-based churn predictor built for ITV using the Cox Proportional Hazards model and the Telco dataset, fit with the Python lifelines package… | ||
| What Failure Taught Me About Building High-Stakes Models | Jake Coltman | 35m |
Models spend over a million dollars a day and initial training optimizes RMSE of revenue per click, which surprisingly rewards a zero‑profit predictor. The… | ||
| How Deliveroo improved the ranking of restaurants | Jonny Brooks-Bartlett | 41m |
Deliveroo creates a point‑wise classification that scores each restaurant’s purchase probability, starts with a simple heuristic, moves to logistic regression… | ||
| A new approach to DataFrames and pipelines | Maarten Breddels, Jovan Veljanoski | 42m |
Vaex claims a working DataFrame that handles billion‑row tables on a laptop by memory‑mapping columnar storage (Apache Arrow / hdf5) and keeping an immutable… | ||
| Prophet at scale to tune & forecast time series at Spotify | Mahan Hosseinzadeh | 39m |
Claims that distributed Prophet reduces runtime from days to about 30 minutes for ~10k series by running models on Spark on GCP with ephemeral clusters… | ||
| On the Path to Causal Inference | Mark Farragher | 41m |
A practical primer that claims causal effects can be estimated without an RCT by drawing DAGs, identifying backdoor paths and controlling the minimal… | ||
| Interpretable AI or How I Learned to Stop Worrying and Trust AI | Ajay Thampi | 34m |
Demonstrates concrete techniques: uses partial dependence plots to expose feature interactions and sampling gaps in a student-grades model, applies SHAP values… | ||
| Because You Can Run, You Can't Hide: Some Musings on API Design | James Powell | 38m |
Claim: Python’s runtime makes true information hiding infeasible. Demonstrations use inspect.getsource and disassembly to copy functions and swap… | ||
| Understanding exchange network dynamics with Python | Omer Yuksel | 31m |
Builds a mixture latency model encoding queueing, bursts and reaction windows, fits it with pymc3 (variational inference) alongside small supervised pieces… | ||
| Kafka in Finance: Over 1 Billion messages a day | Matthew Hertz, Alla Maher | 40m |
1–2 billion market‑data ticks per day arrives from multiple vendors into a real‑time bus then into Kafka as durable replayable buffer partitioned by hashed… | ||
| Productionising Data Science at Scale | Trevor Sidery, Guillermo Barquero | 39m |
Tesco describes production systems that run tens of thousands of forecasts weekly, scale to ~3,000 stores and use Python UDF per row on the cluster with… | ||
| Image search without image captions | Harrison Pim | 26m |
Wellcome Collection faces ~10,000 images a day (~350,000 a month) with sparse captions, so a pipeline maps image feature vectors (ResNet/VGG-style embeddings)… | ||
| Quantamental Investing - the future of finance | Fabian Krause | 33m |
Claims that blending fundamental research with quantitative signals improves discretionary performance and reduces behavioral bias, backed by concrete systems… | ||
| A Primer (or Refresher) On Linear Algebra for Data Science | Ruben van de Geer | 32m |
Claims that linear algebra is the foundation of many ML methods and shows concrete demos: matrix×vector as weighted column sums, a 90° rotation applied… | ||
| Testing and Validating MLMs | Christopher Samiullah, Soledad Galli | 31m |
Reproducibility between research and production is the stated hard claim and the recommended guardrail: reproduce past predictions, then run a shadow… | ||
| The unreasonable effectiveness of feature hashing | Gianluca Campanella | 36m |
Feature hashing maps high‑cardinality categorical and bag‑of‑words features into a fixed, sparse vector with a cheap hash and modulo step, uses a sign bit… | ||
| How to Constrain Artificial Stupidity | Vincent D Warmerdam | 37m |
Claims that adding constraints cuts “artificial stupidity”: predict less (add a do not predict threshold and outlier check via a Gaussian mixture model)… | ||
| Computer Vision and NLP for Multi-Task Fashion Attrb. Modeling | Michael Sugimura | 34m |
Claims an operational pipeline: an ensemble of custom multi-task CNNs and a fine-tuned Google’s Bert in Pytorch that tags millions of products for Shoprunner… | ||
| Training intelligent game agents using deep reinforcement learning | Imran Rashid | 01h 17m |
Claims concrete outcomes: runnable notebooks show how deep reinforcement learning (DRL) agents train with python, using TensorFlow layers, a policy-gradient… | ||
| Understanding of distributed processing in Python | Chie Hayashida | 34m |
Clear comparisons and concrete recommendations: threading hits the GIL blocks CPU threads limit so it’s only good for I/O, multiprocessing works for CPU-bound… | ||
| Weak supervision: a new paradigm for unreliable data | Eddie Bell | 33m |
Claims a practical pipeline: combine many cheap labeling functions into a generative model (using Snorkel) to produce probabilistic labels, then train a… | ||
| How am I going to deal with all of these cats | Liam Kirwin | 42m |
Practical comparisons and concrete tools appear: CatBoost ordered target encoding uses randomized running means to avoid leakage and reduce overfitting, GBM… | ||
| Data science and Internet of Things on the edge | Mario Bonamigo | 35m |
A tiny leak sensor is built and deployed: two temperature sensors on a pipe infer flow from a drop, run fully offline on 32 kilobytes RAM and two AAA… | ||
| A Deep Dive into NLP with PyTorch | Jeffrey Hsu, Susannah Klanecek | 01h 40m |
Hands-on workshop that shows exactly what gets built and how: PyTorch notebooks on GitHub and Colab GPU runtimes, a bag‑of‑words classifier (vocab ~3,000… | ||
| From healthcare.hospital import datascience | Pavlos Papaconstadopoulos, Patrick Gonzalez | 40m |
Project applies anomaly detection to organ contours in head and neck radiotherapy, claiming a pragmatic isolation forest + feature engineering approach… | ||
| Safer Cycling with an Edge TPU watching your back | Zack Akil | 39m |
Build uses a COCO‑trained MobileNet SSD v1 exported to TF Lite, 8-bit quantized TF Lite, then compiled for an Edge TPU USB accelerator attached to a Raspberry… | ||
| Just ask: designing intent-driven algos | Anna Schneider | 32m |
Claims that capturing real-time intent closes the gap between a user’s general profile and what they want right now and demonstrates this by rebuilding… | ||
| Knowledge graphs --enter- the Hype Cycle | George Cushen | 40m |
Claims a working fashion knowledge graph for Farfetch that links products to taxonomy nodes and powers search, recommendations, ranking and personalization… | ||
| Data Science Frameworks + Managed Services | Jon Tutcher | 36m |
Build for simplicity not novelty: a BBC mobile recommendation app takes 18 months to ship after adopting 17 components across GCP, gRPC, Kubernetes and… | ||
| Quantifying uncertainty in Machine Learning predictions | Maria Navarro | 36m |
Claim: conformal prediction gives guaranteed error bounds (1 − ε) with almost no assumptions and works with any model, few assumptions. Method: split… | ||
| Deep Learning and Time Series Forecasting for Smarter Energy | Igor Gotlibovych | 40m |
A production forecasting pipeline built on pandas scikit tensorflow ingests ~77,000 smart meters to train shared neural models with household embeddings… | ||
| Day 1: Lightning Talks | 01h 07m | |
A rapid sequence of concrete demos and claims: a tour of the NumFOCUS project directory and the difference between sponsored vs affiliated projects; a… | ||
| Day 2: Lightning Talks | 39m | |
Grab concrete takeaways: an invitation to submit to Budapest PyData and a community project connecting activists and technologists; a Python time-series… | ||
| nothing matches | ||