INST0098 Chap.7 Researching Social Media Data
Researching Social Media Data
Researching Social Media Data separates questions about platforms from questions that use platform traces to study wider social issues. Access routes, ranking, moderation, deletion and visibility settings shape the observable corpus.
Automated classification can organise scale, but validation and qualitative context remain necessary, especially for irony, quotation, multilingual expression and community specific language.
Write two versions of the question before collection. The platform version asks how features, governance or visibility shape activity. The social version uses traces to investigate a phenomenon beyond the platform.
Mark which parts of each version the available fields can address. This prevents account activity from becoming a substitute for people or visible engagement from becoming a complete measure of attention.
Create a provenance record containing the interface or API, query, dates, filters, returned fields, rate limits and transformation steps. Note private, deleted and inaccessible material as known absences.
Preserve evidence of platform changes because a new field definition or ranking system can create an apparent trend. When measurement changes, mark the break and avoid a behavioural explanation until equivalence is demonstrated.
Construct a collection comparison before committing to one route.
Run the same small query through the visible interface and the programmatic endpoint, then compare ordering, fields, duplicates and inaccessible items. Record which route matches the research purpose and what each systematically omits. If neither offers a stable population frame, describe the corpus as a trace sample rather than implying comprehensive coverage.
Preserve documentation of the access conditions where policy permits.
Validate automated classification on the actual corpus. Select common, rare and uncertain examples rather than only easy cases. Define human coding rules and record disagreement. Inspect errors across language, topic and community position because an acceptable overall score can conceal concentrated harm.
If a category performs unevenly, narrow the use or change the claim instead of presenting the aggregate as neutral.
Contextual reading should identify the target of emotion or criticism, quoted speech, irony and conversational sequence. A negative label applied to grief differs from criticism of an institution. A repost may endorse, mock or archive the original.
Combine pattern detection with close examination of cases that determine interpretation. Keep ambiguity visible when the trace does not resolve it.
Ethical review goes beyond removing usernames. Searchable quotations, screenshots and network positions can reidentify people. Consider expected audience, vulnerability, aggregation and possible harm alongside technical public status.
Decide whether paraphrase protects participants without destroying the linguistic feature under study. Reporting should explain these tradeoffs and limit conclusions to the collected traces and their platform conditions.
Recheck provenance before reporting because platform documentation may change during analysis. Date the methodological account and separate observed behaviour from later platform descriptions.
What this chapter covers
- 01
Platform questions
- 02
Trace data
- 03
Collection provenance
- 04
Automated classification
- 05
Contextual validation
- 06
Search based identification
Worked example · free
Audit a sentiment surge
- 2Check query, fields, access route and collection window for change.
- 2Validate labels on positive, negative and uncertain examples.
- 2Distinguish criticism targets and limit the population claim.
Key terms
- Platform trace
- An observable record shaped by user action and platform systems.
- API provenance
- Documentation of the interface, query, fields and date used for collection.
- Search identification
- Reidentification risk created when a quotation can be found through search.
Researching Social Media Data FAQ
Why is a social media dataset not transparent?
The platform selects visibility through interface design, ranking, moderation and access rules. Private, deleted or down ranked activity may be absent even when it matters to the research question.
Can automated sentiment replace close reading?
It can help organise a large corpus, but meanings such as irony, quotation and target often require contextual validation. The model should be tested on the actual material being analysed.
Is public content always ethically simple to reuse?
No. Formal visibility does not settle user expectation, vulnerability, quotation risk or contextual harm. Ethical reasoning should address the particular community, research purpose and proposed presentation.
Assessment move
Create a provenance sheet before analysing any platform corpus. Record the platform, access route, query, collection dates, visible fields and known absences. Write two versions of the research question: one about platform behaviour and one about the wider issue. Decide which the data can support. Sample examples from every automated category and from uncertain cases, then annotate humour, quotation, target and context.
Compare the machine label with the human reading and revise either the codebook or the claim. Finally, test whether quoted text can locate an account and decide whether paraphrase or omission better protects participants.
Working through Researching Social Media Data in INST0098? Sia is AskSia’s AI Arts and Humanities tutor — ask any INST0098 Researching Social Media Data question and get a clear, step-by-step explanation grounded in how INST0098 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.