ML Problems from CS 441 Students

CS 441 - April 27, 2023

[Problem] by [name]

1. What you are solving and why – use visual examples if possible
2. What are the main challenges (ML and otherwise)
3. Specific questions that you would like addressed/discussed in lecture

​

This can be in 1-3 slides. Please be present in class on April 27 to talk through this and answer clarification questions. Practice explaining – it should take about 5 minutes for the initial explanation

Template: do not change – copy and paste

Identifying CRACKS in concrete bridges with Radar sensing and Machine Learning

A radar sends EM wave and receives a reflected signal that contains information. →

Is this signal from part 1, 2, or 3?

Time (ns)

by Ishfaq Aziz

Each scan-signal is a datapoint which correspond to one class

​

Problem:

• With a given scan-signal, can we predict which class it refers to?

​

Dataset:

• Real data from 5 different bridges with labelled ground truth

​

• Total data = 500,000 signals
• Class-1 : 71 %
• Class-2 : 24 %
• Class-3 : 05 %

→ Imbalanced dataset

Target:

Train on data from 4 bridges and test on data from 5th (zero-shot)

Even after these, models are biased towards the major class (Class-1) →

Models tried:

• Random Forest & Logistic regression with features (Freqs., PCA, etc.) as input.
• 1-D CNN with raw signal as input.

​

Techniques tried for class balancing:

• Oversampling minor class
• Undersampling major class
• Weighted loss function

Input (signal/features)

Output

(class probabilities)

Ideas from class

More ideas

• Amplitude and maybe phase of frequencies over time seems most relevant – use 1D Discrete Cosine Transform (DCT) as the raw representation
• Normalize for each bridge
• Subtract median features from raw features
• Represent the probability of seeing a set of DCT values at a particular time
• Reformulate as detection of depth of crack
• Clustering or mixture modeling may help

Recruiting Currently

Post a job

Manually filter 600-800 applicants

Interview top candidates

Extend job offer

1

2

3

4

The Solution

Hatch is an AI recruitment platform that allows recruiters to rank hundreds

of applicants in seconds.

Link Search: candidate github, linkedin, personal portfolio

What have we considered: What we are exploring:

• Using the Word2Vec model with either SkipGram or CBOW.
• Which is the better choice for resume parsing?
• Then apply the model to the resumes and job descriptions:
• We do we scale the resultant vectors equally. How?
• And apply cosine similarity between the “job vector” and “student vector” to rank the resulting resume vectors:
• Is cosine similarity the best similarity measure? What other comparison metrics can we use?

• How do we build the dataset? We have some beta testers (startup founders, recruiters), but what sort of data should we collect from them? (ranked resumes as truth labels?)
• Are there any foundational models we can start with comparable runtimes and smaller scale?
• How do we eliminate recruiter bias?
• Anything we’re missing when considering this problem!

Ideas from class

More ideas

• Sounds like you have a resume parser, but just in case https://github.com/hxu296/nlp-resume-parser
• Once you know what works, you can use it directly or to inform an approach
• BERT may be good starting point for representation: inputs are json for job description and json for resume
• Building dataset – very tricky.
• One approach is to build a useful non-automated approach that you provide for free and gather ratings of candidate quality and job fit
• Can use RL during deployment
• May want to maintain both general model and client-specific model
• Can predict whether the person is selected for interview, offered, hired – independent recruiters can offer a lot of data
• Bias – consider providing separate scores for different subgroups, various ideas in Model Cards are relevant

​

Vessel Route Prediction by Marcus Kornmann

https://www.vesselfinder.com/

• Automated Information System (AIS)
• Vessels send dynamic and static information in fixed intervals
• Position, vessel specifications (length, width, …), status

Vessel Route Prediction

Goal

• Detect areas of high traffic
• Predict the route of a vessel given its previous movement
• Use information for route planning for autonomous sailing boat

Vessel Route Prediction

Challenges

• Size of dataset: > 6 million data points per day
• How to efficiently read large datasets (even using data for a specific region requires reading all the data once)?
• Very unreliable data:
• Some of the information is added manually => high error quotes (e.g. ship length / width)
• Some specifications are ambiguous (e.g. status)
• Outliers (unreasonable) values for vessel movement

Approach

• Clean data, divide tracks into subtracks, perform line-reduction (define tracks by turning points), perform clustering

Ideas from class

More ideas

• Track flow of traffic instead of points?
• Model density at each tile, and probability of transitioning between tiles
• Potentially, consider longer history, e.g. next tile given previous three tiles
• ​

​

Improving Crystallizer Effluent Particle Size Distribution Predictions with Hybrid Mechanistic/ML Process Models

​

by Sam Aguiar

Can we combine mechanistic models with ML to help better predict changes to crystallizer effluent under non-ideal conditions?

​

Available Data

Currently accounted for in Model

•Ion concentrations

•Particle Size Distribution in reactor

​

Partially included in Model

•Flow rates or Mixing Conditions

•Temperature

​

Not Included in Model

•Organic additives

•Solids content

•Coprecipitation

•Fluid dynamics

How can we implement a Hybrid Model?

Ideas from class

More ideas

• May have some relationship to using RL to solve for kinematics e.g. this
• ​

​

Coming up

• Last class on Tuesday
• Summarizing key ideas from semester
• Where to learn more
• Future of machine learning

​

• Final project due next Wed

​

• Final exam on May 9