# Research and project ideas

This folder stores all the different project ideas I have for potential research. I prefer if you could come up with your own question though.

## Ideas

#### Model for Liebig’s Barrel

In agriculture, the belief is that the shortest resource will dictate the growth of the crop. Is there a way to recover the cause or threshold for each resource using statistical learning?

#### Approximate Bayesian Computation

Besides avoiding the likelihood calculation, ABC is more intuitive to scientists. Can ABC methods be more robust/correct when the model is incorrect?

• Use ABC for amazon review problem: one product with 10 reviews but 5 stars, one product with 100 reviews but 4.5 stars, which one is better?
• Conjugate priors are not intuitive.
• Approximate Frequetist Computation?
• Is replacing the likelihood in MLE with summary statistics just a peusdo-likelihood? What should we expect?
• Should we teach confidence intervals as “candidate intervals” instead?

#### Bayesian Optimization Numerical issue

We know that the joint predictive distribution can be expressed as the product of a conditional distributions. This allows us to sample sequentially without pre-determining the sample locations.

$P(Y_1, \dots, Y_N | Y) = P(Y_1|Y)P(Y_2|Y_1, Y)\dots P(Y_N|Y, Y_1, \dots, Y_{N-1})$

The problem with this, however, is that the conditional distribution is numerically unstable. Are there ways to mitigate this?

#### Data anti-trust as an idea

Can we quantify when someone has “too much” data on a certain topic? How would we even go about doing this? Legal definitions for adverse impact is defined for discrimination, anti-trust also has similar metrics but is more applicable to non-data businesses.

#### Connecting research farm data with grower farm data

It’s common that plant research is done in highly controlled environments but farmer’s data is often much noisier. It’s not obvious how to connect the 2 different sets of data together.

#### WLS in non-linear cases producing bias?

see physics-fun/ for the details. Weighted least squares can produce very biased predictions if the data is highly non-linear and non-Gaussian, why is this given regression can be phrased as a weighted average.

#### Are there usecases for an algorithm that gets “bored”?

Recommendation algorithms dwell on repeated recommendations and can discourage users to continue their subscription. Is there a way to modify the objective function such that algorithms will constantly adapt? Do we need to model boredom?

#### Can we train a linear model without using all the data?

Sufficient statistic, differential privacy, core sets, data sketching, Nyström approximations

#### Data science with dance

Can we detect emotion from people’s movement?

#### Poverty metric

This is currently done by performing PCA on “assets owned”. This method isn’t necessarily robust and is calibrated against consumption.

#### Help students find professors to work with

Explore the citation graphs on professor’s recent publications for students to see.

#### Multi-armed bandit vs A/B testing

Tech companies often start off with A/B testing then start to wonder about replacing it with the multi-armed bandit. The problem is these methods are doing 2 different things: one is effect estimation and the other is optimization.

#### Training on “what you won’t like” instead of “what you will like”

Can we train algorithms only with negative labels, e.g. how babies communicate?

#### Data quality as a research topic

As data pipelines get more complex and models ingest output from each other, data quality goes beyond uncertainty propagation. There is also a human aspect to data quality where earning trust from people who will “take whatever positive result but question all negative results you produce” could use some frameworks.