teaching/enseignement
Machine Learning
See also Course material, LABS
Projects/datasets examples and ideas
BIOLOGY/HEALTH SCIENCES
Despite the availability of various diagnostic tests for inflammatory bowel diseases (IBD), misdiagnosis of IBD occurs frequently, and thus, there is a clinical need to further improve the diagnosis of IBD. As gut dysbiosis is reported in patients with IBD, we hypothesized that supervised machine learning (ML) could be used to analyze gut microbiome data for predictive diagnostics of IBD.
Data are available here, species counts are available here and an associated paper here.
DNA sequences classification [classification, deep learning]
In recent years, a deep learning model called convolutional neural network with an ability of ex-
tracting features of high-level abstraction from minimum preprocessing data has been widely
used. In this project, we propose to use deep learning algorithm for classifying DNA sequences using the con-
volutional neural network while considering these sequences as text data. One-hot vec-
tors will be applied to represent sequences as input to the model; therefore, it conserves the essential position
information of each nucleotide in sequences.
Data can be download here a paper describing some ideas is available here.
A paper with examples can be found here. Data can be downloaded here in the SMILE format. Preprocessing is required.
Eukaryotic protein subcellular localization [classification, deeplearning]
Purity of tumor [regression]
Blood cell images classification [classification, deep learning]:
The diagnosis of blood-based diseases often involves identifying and characterizing patient blood samples. Automated methods to detect and classify blood cell subtypes have important medical applications.
The dataset contains 12,500 augmented images of blood cells (JPEG) with accompanying cell type labels (CSV). There are approximately 3,000 images for each of 4 different cell types. The cell types are Eosinophil, Lymphocyte, Monocyte, and Neutrophil. This dataset is accompanied by an additional dataset containing the original 410 images (pre-augmentation).
Data are available from the Kaggle website here. More details and some ideas to solve the classification problem can be read here.
Breast cancer, images classification [classification, deep learning]:
As the most common malignancy among women worldwide, breast cancer represented roughly 23.8% of all new cancer diagnoses in 2022 [1]. Although X-ray mammography remains the clinical gold standard for early detection, image interpretation is highly specialized and complex. Even with the support of computer-aided detection and diagnosis (CAD) systems, experienced radiologists still face notable error rates. Data are available from the Kaggle website here. Several studies are based on this dataset, see for instance the seminal one
Music and mental health [classification, regression]
Music therapy, or MT, is the use of music to improve an individual's stress, mood, and overall mental health. MT is also recognized as an evidence-based practice, using music as a catalyst for "happy" hormones such as oxytocin. However, MT employs a wide range of different genres, varying from one organization to the next. The MxMH dataset aims to identify what, if any, correlations exist between an individual's music taste and their self-reported mental health. Ideally, these findings could contribute to a more informed application of MT or simply provide interesting sights about the mind.
A paper is available here
Influenza outbreak event prediction via Twitter [classification]
By identifying influenza-related tweets, the goal is to forecast the spatiotemporal patterns of influenza outbreaks for different locations and dates.
The data is from the United States. The data comes from different states under different weeks. For each week, the task is to predict whether or not there is an influenza outbreak on the next date. More specifically, for influenza activity, there are four levels of flu activities from minimal to high according to CDC Flu Activity Map. An influenza outbreak occurrence is indicated if the activity level is high.
Data and paper can be download here.
Down syndrome (DS) is a chromosomal abnormality (trisomy of human chromosome 21) associated with intellectual disability and affecting approximately one in 1000 live births worldwide. The overexpression of genes encoded by the extra copy of a normal chromosome in DS is believed to be sufficient to perturb normal pathways and normal responses to stimulation, causing learning and memory deficits. The questions are to detect/explain the treatment and genotype differences.
References to past uses of the data can be found here (and references therein). The data, their description and an R code to preprocess the data are available here.
In this project, kernel representations are used. These concepts are discussed in the last part of the course. But, help will be provided if you choose to work on this subject.
The aim of our study was to compare the results of unrelated donor (UD) peripheral blood stem cell transplantation versus UD bone marrow transplantation and to analyze the impact of infused CD34+ and CD3+ cell doses on survival and incidence of severe graft-versus-host disease (GVHD) in 187 children who underwent UD hematopoietic cell transplantation with the use of in vivo T cell depletion (antithymocyte globulin or CAMPATH-1H).
More details and the dataset are available here.
Lymphoma [classification]
The lymphoma dataset (Shipp et al., 2002) consists of 7129
gene expression levels from 77 lymphomas. The 77 samples
are divided into 58 diffuse large B-cell lymphomas (DLBCL)
and 19 follicular lymphomas (FL). The data can be found
at https://github.com/ramhiser/datamicroarray/blob/master/data/shipp.RData and
the reference paper here.
Colon cancer tumor [classification]
The colon dataset is from the microarray experiment of
colon tissue samples of Alon et al. (1999). It contains
the expression level of 2000 genes for 40 tumors and 22
normal colon tissues. The data can be freely downloaded
from http://microarray.princeton.edu/oncology/affydata/index.html
and the reference paper here.
however, it may also be related to previous simian virus 40 (SV40) infection and quite possible for genetic predisposition.
Molecular mechanisms can also be implicated in the development of mesothelioma.
The question here is to propose machine learning algorithms to predict the diagnostic.
Predicting wildfire spread is critical for land management and disaster preparedness. ‘Next Day Wildfire Spread,’ is a curated, large-scale, multivariate data set of historical wildfires aggregating nearly a decade of remote- sensing data across the United States. In contrast to existing fire data sets based on Earth observation satellites, our data set combines 2D fire data with multiple explanatory variables (e.g., topography, vegetation, weather, drought index, population density) aligned over 2D regions, providing a feature-rich data set for machine learning. A paper describing an analysis of the data set is available here. The dataset is available here.
Can we predict the structure of soil microbial communities using environmental data as effectively as we do for macro-organisms? It’s a tough challenge: high microbial diversity and scarce benchmark data make soil ecosystems notoriously difficult to model.
A paper describing an analysis of the data set is available here. The dataset is available here.
Species richness is the core of biodiversity-ecosystem functioning (BEF) research. Nevertheless, it is difficult to accurately predict changes in plant species richness under different climate scenarios, especially in alpine biomes. In this project we propose to identify the most critical driver of species richness distribution.
A paper describing an analysis of the data set is available here. The dataset is available here.
Accurate natural resource inventory information is vital to any private, state, or federal land management agency. Forest cover type is one of the most basic characteristics recorded in such inventories. Generally, cover type data is either directly recorded by field personnel or estimated from remotely sensed data. Both of these techniques may be prohibitively time consuming and/or costly in some situations. Furthermore, an agency may find it useful to have inventory information for adjoining lands that are not directly under its control, where it is often economically or legally impossible to collect inventory data. Predictive models provide an alternative method for obtaining such data. The question is then to propose machine learning algorithms to predict the cover type from cartographic variables only (no remotely sensed data).
Impact of Bedrock Fractures on River Erosion [classification, deeplearning]
In bedrock-dominated upland terrains, local heterogeneity in the erodability of rock masses is a critical but under-explored factor constraining sediment erosion, mobilisation and transport. In this project, we propose to analyze original data from laboratory experiments. The experience is detailled here. It results in a large amount of images presenting areas of plucking. The objective is to propose a deep learning algorithm to automatically discover the plucking areas. Data have to be asked to me.
Clouds type detection [classification,
deep learning]
Clouds are a major challenge for passive satellite
imaging, and daily cloud cover and rain showers in the
Amazon basin can significantly complicate monitoring in
the area. For this reason, we have chosen to include a
cloud cover label for each chip. These labels closely
mirror what one would see in a local weather forecast:
clear, partly cloudy, cloudy, and haze. For our purposes,
haze is defined as any chip where atmospheric clouds are
visible but they are not so opaque as to obscure the
ground. Clear scenes show no evidence of clouds, and
partly cloudy scenes can show opaque cloud cover over any
portion of the image. Cloudy images have 90% of the chip
obscured by opaque cloud cover. The dataset to be used
gathers satellite images of Amazon and labels (clear,
partly cloudy, cloudy, and haze) ; it can be downloaded
from the Kaggle website here.
This talk
describes an analysis of this dataset.
♣: if you decide to work on this project, the dataset is
available in a google drive (link).
An example of code for starting is proposed under this link. Thank you to save a copy in
order that the code will not be changed.
Butterflies are abundant species on the earth, and the task of identification of butterflies is complex. How toapply image processing methods to automatic identification of butterfly species is a hot issue in current research.
Birdcall identification
[classification, deeplearning, sound preprocessing]
Do you hear the birds chirping outside your window? Over
10,000 bird species occur in the world, and they can be
found in nearly every environment, from untouched
rainforests to suburbs and even cities. Birds play an
essential role in nature. They are high up in the food
chain and integrate changes occurring at lower levels. As
such, birds are excellent indicators of deteriorating
habitat quality and environmental pollution. However, it
is often easier to hear birds than see them. With proper
sound detection and classification, researchers could
automatically intuit factors about an area’s quality of
life-based on a changing bird population. it can be
downloaded from the Kaggle website here.
The challenge is to build machine learning
algorithm(s) to predict the bird species from audio
records. Deeplearning is only one of the possible
solution.
You will find lectures about deep learning for audio here.
Road A project concerned part of the planned A1 motorway section in Pyrzowice; the section is located along the northern border of the Silesian Voivodship and is about 75 km long. The field research involved a strip of land with a width of 500 m on both sides of the proposed project area. Finally, the first project included 80 amphibian breeding sites. Data are available or a second road.
Predicting CO2 Emissions [regression]
The ability to accurately monitor
carbon emissions is a critical step in the fight against
climate change. Precise carbon readings allow
researchers and governments to understand the sources
and patterns of carbon mass output. While Europe and
North America have extensive systems in place to monitor
carbon emissions on the ground, there are few available
in Africa.
The objective is to create a machine learning
models using open-source CO2 emissions data from Sentinel-5P
satellite observations to predict future carbon
emissions.
These solutions may help enable governments, and other
actors to estimate carbon emission levels across Africa,
even in places where on-the-ground monitoring is not
possible.
Material is available here.
Air pollution in large cities produces numerous diseases and even millions of deaths annually according to the World Health
Organization. Pollen exposure is related to allergic diseases, which makes its prediction a valuable tool to assess the risk level to
aeroallergens. However, airborne pollen concentrations are difficult to predict due to the inherent complexity of the relationships
among both biotic and environmental variables.
In this project, you will aim at building Machine Learning models for prediction of presence/absence (or concentration) of pollens in air given meteorological data. This paper can give a first overview of meteo variable that can be usefull for pollen emission prediction. However, you may build/find other variables to improve the models. Please ask the teacher to get the data.
Weather data come from the ECA dataset and are recorded at Luxembourg airport.
A paper with an example here and a paper with some description of the data here. The datasets are available below for the forecast horizons 15h and 27h with a README file.
Prediction of ozone pollution in Houston.
[classification, rare event]
Accurate ozone alert forecasting systems are necessary to
issue warnings to the public before the ozone reaches a
dangerous level. However, little is known on exactly what
features are important in ozone production and how they
actually interact in the formation of ozone. This provides
wonderful opportunities for machine learning.
One of the reference paper here
and a short description of the original dataset here
which contains 72 meteorological variables and a version
with no missing values below (Ozone_imputed.csv).
Communities and Crime in the US [regression]
A description of the data can be found here and the data can be download here.
The dataset includes missing data.