| theme | gaia |
|---|---|
| _class | lead |
| paginate | true |
| backgroundColor | |
| footer | Computational Thinking and Social Science | :copyright: Matti Nelimarkka | 2023 | Sage Publishing |
| marp | true |
- Be able to describe and implement algorithms to transform data from surveys or text documents
- Describe the elements that distinguish rule-based, supervised, and unsupervised analysis approaches
- Understand suitable areas of application for each of them
- Describe at a conceptual level how each of them
- Know which algorithmic tools are used in each of them
- Understand the practices of conducting data science
- Helps to classify or predict values based on data
- Works both on quantitative and qualitative data
- Suggests new methods and practices for data analysis
- Data science is not a revolution, but an extension
-
Data modelling culture: finite hand-chosen and theory-drive predictive variables are put into the model, and they are used as building blocks for the statistical model.
-
Algorithmic modelling culture: the predictive variables are not a priority, but focus is shifted to produce the highest accuracy model using more variables than most in data modelling culture would like to
-
Is algorithmic modelling culture atheorethic? Data are always contextual, thus social science imagination is required.
- Collected strictly for research purposes
- Emerging as a by-product of other activities
- a representative body of survey data collected for research purposes
- data from national statistics offices or similar resources
- transcripts of interviews conducted for research purposes
- digitalised discussions - for example, from parliamentary hearings
- status updates from Twitter
- videos uploaded to TikTok
Data science works with numbers. Therefore, data need to be transformed into a suitable numerical representation, or features.
- Surveys and quantitative data
- Cleaning data (missing values etc.)
- What kind of scales are used and does it impact modelling choices?
Data science works with numbers. Therefore, data need to be transformed into a suitable numerical representation, or features.
- Text as data requires a lot of preprocessing
- Cleaning unnecessarily words and markings
- Reducing the complexity of language
- Often produced as document-term matrices or word embeddings
Data science works with numbers. Therefore, data need to be transformed into a suitable numerical representation, or features.
- Multimedia content requires even more preprocessing
- Extracting features: colours, brightness
- Image recognition systems may extract objects
- Group discussion: Exercise 3.8a
- Individual or group discussion: Exercise 3.1
- Pair- or group discussion: Exercise 3.2
- Researchers manually devise the rules for the data analysis.
- This usually is manual labour with quite an extensive number of
ifandelse-rules. - Good: Researchers know exactly what is the criteria used
- Bad: The criteria is limited by researchers imagination
- Pair discussion: Exercise 3.3
- Transform Code example 3.1 into a flowchart (see Figure 2.2)
- Pair discussion: Do you agree with Code example 3.1? Would you have done it differently?
- Programming exercise: 3.16
- Reproduces an existing classification
- Requires examples from the data
- Requires ground-truth values, or correct classifications for some part of the dataset
- Similar to regression analysis
- Good: Researchers control the classification in use
- Bad: Researchers cannot easily control the impact of variables
- Various algorithms exist: decision trees, random forests, support vector machines, neural networks, ...
- Different algorithms may have different expectations on the data, and may not suite in all cases
- Often researchers try various algorithms and examine what produces the best model performance
- Overfitting is a problem: try train/test split or k-fold cross-validation
- Individual exercise: Exercise 3.9
- Grouping data units based on statistical analysis on similarity and differences
- Does not require human guidance
- Researchers must articulate the meaning for the outcomes
- Similar to exploratory factor analysis or grounded-theory process
- Good: Allows data-driven analysis
- Bad: The interpretation is not always easy
- Various algorithms exist: principal component analysis, self-organising maps, k-means, topic models...
- Algorithms always spit out a solution to the problem: validation of outcomes is critical
- Individually or in pairs: Exercise 3.5
- Group activity: Exercise 3.8b-c
- Homework: Exercise 3.6 or 3.7
- Programming: See suitable programming exercises from chapter.
- What purposes does data science serve in social science research?
- How can non-quantitative data (such as text or images) be transformed into computational data, and what issues does this entail?
- List the similarities that connect rule-based data analysis, supervised, and unsupervised machine-learning algorithms.
- List the elements that distinguish them algorithms.
- What kinds of research and data management practices do supervised and unsupervised machine-learning methods involve?