More graph features, PageRank and hyperparameter tuning | Weekly Report 43

05 Jun 2017 . category: report . Comments
#kaggle #neural networks #meetup

My weekly report from the 29th of May to the 4th of June – productive week, i must say. Running short on sleep!

The Quora challenge continued, as i worked on a handful of features this week – implemented the weighted graph structure for features like weighted neighbors ratio and weighted neighbors contribution. Also used the graph to make a basic PageRank model, i even released a kaggle kernel on it, good response so far!

For the past 3-4 days, i’ve been struggling in building a strong keras pipeline, and tuning the hyperparameters. Have come a long way in learning about the technical details though, it’s been good learning. The neural network has given a decent score, but not better than my xgboost model, currently lingering around the 500 mark on the leaderboard. A couple of days left in the competition, i’ll try a couple of models, probably make a simple stacker and figure out the oversampling piece. Expect a comprehensive blog on QuoraDup soon!

With QuoraDup ending and the bronze medal looking tough, i’ll be focussing on a couple of other competitions now – probably the Mercedes manufacturing one and the indiahacks ml challenge.

At work, i’m back on the UserJourney piece – trying to derive insights from usage patterns. A regular in the day of a Product Manager! Earlier in the week, the analytics db was facing some issues and i had to figure out the backup queries for a day or so – stressful. Attended the June townhall, inspiration as usual!

This weekend, i spent a day with DataKind volunteers working on some cool ML/Data viz projects – it’s a really good community of passionate data guys who come together to solve various real-world problems faced by NGOs. I worked on a project involving “Pollinate Energy” to identify certain communities via proxy data like education, health and transport facilities in those specific areas. While there were a lot of ML challenges(less data, duplicates, sampling biases, unbalanced data), but i found the data collection challenges to be more intersting! Was a good experience, raring to go again!

The Nike Run plan is up and running, went for three runs involving a tempo as well as a long run! Also went for a much-wanted tennis session earlier yesterday – Bangalore Garros ;)

That’s it folks! Later.

---