"Who has better engagement" is the shallow question.
Comparing two media brands by follower count or like totals tells you almost nothing about how they occupy a platform. Wired and The Verge cover overlapping ground, but the shape of the conversation around each — its themes, its tone, its network of voices — is where the real difference lives.
The goal was a brand-intelligence read: not a scoreboard, but a multi-angle comparison of sentiment, topics, network structure, and positioning, each stage building on cleaned, shared data.
Engagement counts the noise. Topics and networks find the signal.
Discipline for a multi-stage pipeline.
One stage per notebook.
Collection, cleaning, analysis, sentiment, topics, network, and perceptual map each live in their own notebook, so any stage can be rerun without re-collecting data.
Raw and processed kept separate.
Raw captures are frozen; processed datasets and JSON summaries are derived — reproducible from the raw layer, never edited in place.
Every figure is an artefact.
Each chart and table writes to
outputs/with a stable index, so the report cites files rather than screenshots.
Seven stages, cleanly chained.
The pipeline runs in order — collect, clean, analyse, score sentiment, model topics, build networks, map positioning — each reading the previous stage's processed output.
The most interesting stage is the network work: keyword co-occurrence graphs plus centrality scoring surface the micro-influencers — smaller accounts that sit structurally central to a brand's conversation and would be cheap, high-leverage partners.
Different themes, different tones, different maps.
Stacking the stages produced a genuinely multi-dimensional comparison: engagement distributions and posting-time heatmaps, sentiment splits over a rolling window, distinct LDA topic clusters per brand, and finally a perceptual map placing the two brands against each other on derived axes.
The payoff of the layered approach is that no single chart is asked to carry the verdict — the perceptual map is only trustworthy because the sentiment and topic stages beneath it are visible and reproducible.
Scores that point at people.
The deliverable ends with recommendations, not just descriptions: per-brand micro-influencer shortlists ranked by network centrality, and a positioning read from the perceptual map. That's the shift from social analytics to social intelligence — output a marketing team could act on tomorrow.
Three things, next time.
Use a transformer sentiment model.
Lexicon sentiment struggles with sarcasm and tech jargon; a fine-tuned transformer would read tone on this domain more reliably.
Track topics over time.
Dynamic topic modelling would show how each brand's themes shift week to week, not just their static mix.
Orchestrate the notebooks.
Seven notebooks run by hand invites drift; a small pipeline runner would guarantee the stages execute in the right order every time.