Knowledge Discovery and Data Mining 2 (2021/22)

Course Organization

Introduction to the course and the administrative aspects.

Ensemble Methods

Combination of multiple learning algorithms/models/hypothesis

Each learning algorithm might have different strength and weaknesses. The idea of an ensemble is to combine the (weak) learners (e.g., base classifiers) into a combination that eliminates some of the weaknesses and combines some of the strengths.

Today, ensemble algorithms like Random Forests or Gradient Boosting are goto-methods for many data science tasks.

Time Series Data

Time series data requires specific preprocessing and analysis.

Anomalies in Data

Outliers/Anomalies/Surprise/Noise - all the same, or different?

Graph Data Science

Graphs can model complex systems, such as, social, biological and web networks and even molecules and meshes. Graphs require specific processing and analysis compared to other data structures, i.e., tabular, image, or time series data.

Causal Data Science

How does causality help in Data Science?

Privacy-Preserving and Fairness in Data Science

Confidentiality and fairness in Data Science. Introduction to main concepts, including k-Anonymity, Differential Privacy and Federated Learning.

Bias & Assumptions in Data Science

What assumptions do we have to make to effectively conduct data science? Is our data biased, or our algorithms? And why? And what can we do about that?

Task: Take part in an scientific challenge (shared task), or an Kaggle challenge of your choice. More details about the data team and how to get in touch can be found on the Data-Team Homepage.

Task: Given some observations (data), find the instances that do not conform to the remainder of the observations.

Approach: Select a dataset and setting. Analyze the dataset, search for appropriate anomaly detectors, implement & run them on the dataset, and report their performance.

Suggested datasets:

• ODDS
  Large collection of outlier detection datasets with ground truth: ODDS
• DAMI
  Datasets used in scientific publications (complete with performance results): DAMI

Advanced #1: Implement your own anomaly detection algorithm and compare its performance against the baseline.

Advanced #2: Work with time series datasets (if interested, please contact instructors).

Task: Given a dataset, which contains some sensitive information, the dataset should be transformed into a representation (e.g., a modified version of that dataset), which no longer contains the sensitive information.

The type of the sensitive information is defined beforehand and could either be regarding the membership (e.g., if a certain person is part of the dataset), or some attribute (e.g., the income).

Task: Propose a dataset, which may contain information that might lead to unfair results, unless special methods are applied. It should be clear that in a default data science pipeline some instances (e.g., a sub-population) would be treated unfair.

It is expected that first a definition of fairness is introduced for the dataset (i.e., what do we consider as fair in this context).

Task: Identification of causality relationship directly from the data (i.e., cause, effect). This is an advanced data science task with big potential.

Data-Set: Data from a past challenge (complete with prior research): Causality Challenge #1.

Papers:

• Runge, J., Petoukhov, V., Donges, J. F., Hlinka, J., Jajcay, N., Vejmelka, M., ... & Kurths, J. (2015). Identifying causal gateways and mediators in complex spatio-temporal systems. Nature communications, 6, 8502.

Task: Collect datasets (e.g., open-governmental) datasets, analyse these and assess the suitability of these datasets for a number of application scenarios.

Approach: Research datasets, assess their key characteristics, apply data science methods to assess their usefulness.

Advanced: Build a database (or similar) to allow to collect and update the relevant key parameters of each dataset.

Task: In data science we often we want to prevent (or predict) failures. For example we want to avoid accidents to happen due to fatigue of material. Therefore, it is necessary to derive a health index from the data that predicts e.g. the remaining life span of parts.

Data-Set: A suggestion for starting point: NASA Prognostics Data Repository, especially the dataset mentioned in the paper below.

Papers:

• Arias Chao, M., Kulkarni, C., Goebel, K., & Fink, O. (2021). Aircraft Engine Run-to-Failure Dataset under Real Flight Conditions for Prognostics and Diagnostics. Data, 6(1), 5.

Task: Automatic tagging of resources, using either an unsupervised or a supervised approach. The goal is to apply tags to an unseen resource.

Approach: For this project the approach may vary widely. A supervised approach requires a training dataset and may include classification algorithms. Unsupervised approaches may either look at a set of resources or a single resource at a time.

Suggested data-sets:

• Stack Exchange
  One example for tagged data are the stack exchange pages, which can be downloaded here: Stack Exchange Dump
• Last.fm
  The Million Song Dataset, which contains tracks, similar tracks as well as tags. The dataset is already split into a training and testing dataset and can be accessed here: Last.fm Dataset

Advanced: Implement an unsupervised and a supervised approach, and then compare the two approaches. Measure their differences in accuracy as well as discuss their individual strengths and weaknesses.

Task: Compare two machine learning models with each other and explain/interpret what they have learnt. For example, which features have been picked by the respective models.

Approach: Train two separate machine learning models - they should be distinctively different, for example a CNN and a linear model.

Advanced: Compare approaches what are predominantly physical-driven (i.e., the features are derived from expert Knowledge or physical laws) with data-driven (i.e., no explicit feature engineering) approaches.

Task: Given a set of sensor data for multiple streams, e.g. temperature, power consumption. The goal is to predict the future values of these signals, optimally including a confidence range.

Approach: Take the stream of data and remove the last ~10% of the data. Build a prediction algorithm that is able to predict the future values of the streams as accurately as possible and compare against the values you removed.

Suggested data-sets:

• Intel Berkeley Research Lab
  Take for example the data from the sensors of the Intel Berkeley Research Labs, see Stream Data Mining Repository. The data is in a format used by many machine learning frameworks, e.g. Weka.
• Powersupply Stream
  Use the power supply stream from the same data source: Stream Data Mining Repository. Here the challenge is to integrate seasonality into the analysis.
• UCI Repository
  Repository of multiple data-sets, including timeseries.

Advanced: Investigate your prediction algorithm and try to determine under which (controlled) circumstances it makes correct predictions.

Task: Classification of time series data.

Approach: Pick a dataset from the linked repository of time series datasets and try to reproduce (or surpass) the posted performance values.

Suggested data-sets:

• Timeseries Repository
  Web-site of timeseries dataset together with the performance of a number of different algorithms: Welcome to the UEA & UCR Time Series Classification Repository.

Advanced: Analyse how the performance (i.e., classification accuracy) drops, the fewer data is used (fewer parts of the timeseries), to simulate a early classification problem.

Task: Given a set of sensor data for multiple streams, e.g. temperature, power consumption. Apply sequential pattern mining on a time series dataset.

Suggested datasets:

• UCI Repository
  Repository of multiple data-sets, including timeseries.

Advanced: Preprocess the data using the Matrix Profile technique and analyze the effect this has on your results.

Task: Detect the seasonality (= number of observations of the dominant repeating pattern in time series).

Approach: Take the stream of data, build your own season length detection algorithm and compare against an existing algorithm.

Suggested data-sets:

• CRAN Time Series
  Time series dataset, see paper on how to use this dataset: Toller, M., Santos, T. and Kern, R. (2019) ‘SAZED: parameter-free domain-agnostic season length estimation in time series data’, Data Mining and Knowledge Discovery. doi: 10.1007/s10618-019-00645-z.
• NOAA Water Level
  For example the data from the currents NOAA Water Level.

Scenario: A user is starting to search by typing in some words...

Task: The system should automatically suggest word completions, depending on the already entered words.

Suggested data-sets:

• Wikipedia
  The Wikipedia can be downloaded as dump, either as XML or as MySQL Database from the Wikimedia website.
• Europeana
  Instead of using a dataset to retrieve relevant items from, one can directly use a search engine. The Europeana project directly supports a JSON query interface, which can be accessed with an API key: Europeana API Portal

Advanced #1: Provide a estimate of the number of hits for each completion suggestion.

Advanced #2: Provide not only completions, but also similar queries (yielding similar search results), may also include synonyms.

Task: Provide a list of matching resources for a given piece of text. The goal is to produce a ranked list of items relevant to a context.

Use-Case: Consider a user writing a text, for instance a blog post. While typing the user is presented a list of suggested items, which might be relevant or helpful.

Suggested data-sets: Same as previous task.

Advanced: Identify Wikipedia concepts within the written text. For example, if the text contains the word Graz it should be linked to the corresponding Wikipedia page.

Task: Collect a graph dataset, analyze its properties (e.g., attributes, homophily, graph structural properties), and define a prediction task on it, such as, node classification or link prediction.

Example: crawl a network of articles on a certain topic of your choice. Edges can be the web links between the articles. You can use the InfoBox or the first paragraph to create node features/labels.

Task: Select a publicly available graph dataset with a task (e.g., node property prediction, edge property prediction). Analyze the dataset, and provide approaches for solving the respective task.

Tools: network analysis (e.g., with networkx), label/attribute propagation, graph embedding, graph neural networks.

Resources for graph data: PyTorch Geometric, SNAP, OGB (for large scale data).

Task: Select a publicly available non-graph dataset (e.g., text corpus, tabular data, image data, ...) with a certain task (e.g., classification, regression). Find a way to build a graph from this data to improve the performance of traditional models through graph models. Study your built graph by analyzing homophily and find which models are suitable for the task. Train your models to solve the respective task. It's not required that your solution significantly outperforms the traditional solution.

Tools: network analysis (e.g., with networkx), label/attribute propagation, graph embedding, graph neural networks.

For inspiration, see section 5.2. Downstream Tasks in this paper and Text GCN section in this paper.