Starting Our Kaggle Meetup
“Anyone interested in starting a Kaggle meetup?” It was a casual question asked by the organizer of a paper-reading group. A core group of four people said, “Sure!”, although we didn’t have a clear idea about what such a meetup should be.
That was 18 months ago. Since then we have developed a regular meetup series that is regularly attended by 40-60 people. It has given scores of people exposure to hands-on data science. It has also connected numerous startups and established companies with people looking for career opportunities in data science.
Participating in this meetup has been a very rewarding experience for all involved. In this blog post we’ll share what we’ve learned and hope it will encourage others to give it a try.
It is fairly well known that a great way to practice and hone your data science skills is to join a live Kaggle competition. Less well known is that completed competitions have a lot of value for both beginners and experts.
If you are new to data science, Kaggle has “Getting Started” competitions that are an excellent way to take your first steps. But there is also a wealth of material related to the regular competitions. After a competition is completed, the top-ranked competitors are interviewed for the Kaggle blog No Free Hunch. Kaggle participants are incredibly generous with their knowledge and many will post descriptions of their solutions, as well as code and helper scripts. Often the competition data is still available and you can submit your own solutions to see how you rank on the leaderboard.
So on the one hand we have a very rich resource for learning about data science, and on the other there are many people who want to learn data science but aren’t sure how to get started. Our Kaggle meetups have proven to be a good way to bring these elements together. In the following we outline what has worked for us, in the hope that it will be helpful for others who would like to do something similar.
Getting Started: Set Your Goals
It is important to ask yourself why you are doing a Kaggle meetup at all. There are lots of decisions to be made and it is important to have some core principles to refer to.
For us the primary goal is education. The core members that started the meetup were interested in educating themselves and opening it up to anyone else who was looking to do the same.
A secondary but very important goal is networking. Our purpose is not to work on active Kaggle competitions together, but we do give attendees an informal way to find Kaggle team members. Or start a company, or do a hackathon, or whatever. We are also not explicitly trying to be matchmakers between employees and employers, but inevitably it happens, and that’s great.
Organizers (Plural) are Critical
We strongly recommend having at least two organizers for the meetups. (We have two.) You want to have a backup in case someone is away. Also, there is enough work to do (finding a room, posting on Meetup, organizing presenters) that it is a bit much for a single person to fit into their spare time.
After some experimentation, we have ended up with the following format for the meetups.
- We use Meetup.com to manage the scheduling and advertising of the meetings. It’s not perfect, but people who are interested in learning about data science will probably be able to find you there.
- We use Slack to provide a place to discuss competitions and related topics in between the meetups.
- Meet regularly. In our case, it’s every two weeks. That gives people enough time to digest one competition and move on to another one without consuming their lives completely.
- The maximum number of people we have at a meetup (about 60) is limited by our room size. However, this has turned out to be a good number. It’s big enough to ensure there is a lively discussion at each meetup, but small enough that everyone has a chance to meet everyone else and contribute to the meetup.
- Pick a specific competition to talk about at each meetup. This can’t be a live competition, to stay within the Kaggle rules. But a live competition wouldn’t fit what we are trying to do anyway. Instead, stick with the getting started competitions (like Titanic) or completed competitions.
- Have someone do a presentation about the competition. This has turned out to be very important, but we didn’t do it in the beginning. Initially we told everyone to read up on the competition and come prepared to have a roundtable discussion about it. But people were too busy to do that homework, so it wasn’t very successful. Once we started having presenters, it worked better for everyone. And attendance shot up.
- Schedule about an hour for the presentation, including discussion. We don’t chop off a lively discussion because of the clock, but we balance that with being respectful of people’s time and attention span. If you want to keep talking, do it over beer afterward!
- How to choose what competition to talk about? The presenter has the final decision. We encourage people to pick competitions we haven’t covered before, but that is not a hard requirement. We generally prefer more recent competitions to older ones. It is surprising how quickly data science techniques evolve and a competition that is just a few years old can feel pretty ancient.
- Encourage the presenter to make their slides available a day ahead of the meetup. Also encourage them to include links to background information in their slides. This gives everyone a chance to prepare a bit for the meetup, and is particularly appreciated by the beginners.
- At the beginning of the meetup we do a quick round of self-introductions (networking!) and take a couple of minutes for announcements (“We’re hiring!”, “Please join my hackathon”, …).
- We open up the doors half an hour before the meetup starts, to give people a chance to chat.
- After the meetup, we go to a nearby restaurant for food and/or drinks. This has turned out to be a critical part of the success of the series. (Networking again!) Some would argue that it is the best part of the meetup. Among other things it is a good time to ask first-time attendees for feedback about the meetup and what suggestions they have for improvements.
- We have experimented with live streaming and video recording the meetups, but we don’t do it routinely. It adds complexity and can put extra pressure on the presenter. But sometimes the presenter wants to have it recorded to gain more exposure or just have the work they’ve done reach a bigger audience. We would love to hear from others about their experience with or thoughts on streaming and recording. It would be a natural way to broaden our reach.
- Finding a good meeting room is important. Location, location, location: make sure it is accessible by transit, is close to where people work, has a good projector, etc. Look for a sponsor who will make a room available for free. Local colleges, startup hubs, or supportive businesses are the ones to target.
- We don’t charge money for any reason and we don’t seek financial sponsors. Things get exponentially more complicated if money is involved.
We have not had too much trouble finding presenters. That has been a pleasant surprise for us, but here’s why it works:
- The primary goal is education and there is no better way to learn than to present to someone else. So presenters get a personal benefit.
- Reasonably often one of our members has participated in the competition being discussed. They are generally happy to talk about it, no matter what rank they achieved.
It gives the presenter exposure and a chance to practice their presentation skills in front a friendly audience. (So be sure to make the meetup a welcoming and safe place0.)
Sometimes people have to be gently encouraged to be a presenter, but it is worth making the effort to do so. Everyone benefits from having a wide variety of people doing the presentations.
Guidelines for Presenters
We don’t have formal guidelines for presenters, but they often ask for them. Here’s what we say:
- You are not expected to know everything or be an expert about all the techniques that come up. Members of the audience can often fill in the gaps of your knowledge. The presentation should really be more like a discussion. Just be sure to include enough concrete information to generate that discussion. The more questions and answers from the attendees, the better.
- Think of your audience. It is a mix of newcomers, experienced Kagglers and everything in between. Try to ensure that everyone, no matter what their level of knowledge, can get something out of the presentation. You were a noob once. Try to remember what that was like and what kind of presentation you would have wanted to hear.
- Don’t be afraid to explain some basic things in simple language. (Briefly!) You can be sure that many people in the room have imposter syndrome and will appreciate a little spoon feeding.
- However, resist the urge to explain complex concepts like convolutional neural networks (CNNs), XGBoost, etc. There just isn’t time to convey how CNNs work to someone who has never heard of them before. But challenge yourself to come up with an understandable introduction to such concepts by describing when the techniques might be used and why. That is a useful starting point for someone who wants to go deeper. Keep it to a couple of sentences.
- Be sure to take enough time to describe the problem that is being addressed by the competition. This is generally something that can be understood by everyone and really engages the audience.
- Show enough examples of the data so it is clear what the competitors had to work with. And what might have caused problems for them.
Be clear about what competitors are being asked to predict and what the evaluation criteria are. Don’t force your audience to have to ask, “Um, is a bigger number better or worse?”
- Once you’ve done all that level-setting, feel free to describe the solution details in more of a shorthand way that is geared to the more experienced members of the audience.
- You can choose to discuss just one solution and go deep on that, but it’s usually more interesting to hear about a variety of approaches that were tried. Hearing about attempts that didn’t work are informative too!
- It is a good idea to at least skim through the discussion forums. There is a wealth of insight and competition backstory there.
- Look for blog posts on No Free Hunch and elsewhere by the winners. It can make your preparation job much easier!
- As above, post your slides and links a day ahead of the meetup.
Our Kaggle meetup has discussed over 30 different competitions. A list of them is here, and that list includes links to the slides and the occasional videos that we have made. That should give you a good idea of what we do.
The steady turnout of attendees we get is evidence that the format is successful. Those who have attended the meetups have been exposed to a wide range of interesting data sets, challenging and important data science problems, and the latest techniques for solving such problems. Our attendees include students, people who do data science for a living, and many who are exploring a career transition into data science.
Where do we go from here? The Kaggle discussions have whetted people’s appetites to learn more about the techniques commonly used. We have spun out another meetup specifically for deep learning. It is a self-study and discussion group based on the fast.ai course, which incidentally, is structured around one of the Kaggle learner competitions. It was almost instantly oversubscribed and we will likely re-run it in the future, and look to expand the offering to other topics.
In the meantime, we will continue our Kaggle meetups and are glad that there is a steady flow of new competitions for us to discuss. If you’d like to start your own Kaggle meetups, we would be happy to help in any way that we can and look forward to hearing about your experiences. Feel free to contact the author at the address above.
Note from Kaggle: If you’re launching a new Meetup, we’d also love to help reach out to other Kagglers in your area. Get a hold of us at firstname.lastname@example.org.