Earlier work · NLP
Increasing Engagement on National Geographic's Instagram
Predicted Instagram engagement from Google Vision image labels and captions (77% accuracy), then used LDA topics to show that wildlife posts drive the most engagement.
Business question
National Geographic (@natgeo) posts striking photography to one of Instagram’s largest audiences. Which kinds of images drive engagement, and which don’t? The goal was to give NatGeo concrete advice on what to post more and less of.
Data
- Posts. The
instaloaderpackage collected NatGeo posts with their captions, like counts and comment counts. - Image labels. Each image was sent to the Google Cloud Vision API, which returned descriptive labels such as wildlife, mammal, sky or monochrome. That gave about 1,500 labelled images.
Measuring engagement
Likes and comments were each normalised by their maximum, then combined into one score:
engagement = 0.4 × likes (normalised) + 0.6 × comments (normalised)
Comments get the higher weight because writing one takes more effort than tapping like. Posts were then split at the median into high and low engagement.
Predicting engagement
Three logistic regression models were trained on text features (bag-of-words and TF-IDF), each using different inputs:
| Predictors | Accuracy |
|---|---|
| Image labels only | 74.8% |
| Captions only | 74.2% |
| Image labels + captions | 76.6% |
What’s in the picture predicted engagement slightly better than what the caption said, and combining both worked best.
What the images are about: LDA topics
Topic modelling (LDA) on the image labels found four clear themes:
| Topic | Top labels |
|---|---|
| Wildlife | wildlife, animal, mammal, terrestrial, vertebrate |
| Aerial landscape | sky, landscape, natural phenomenon, mountain, tree |
| Water | water, sky, sea, architecture, turtle |
| People | photography, monochrome, black, people |
Posts were grouped by engagement (bottom 25%, middle 50%, top 25%) and the average weight of each topic was compared across groups:
- Wildlife had the highest weight among the most engaging posts.
- People, especially monochrome portraits, had the highest weight among the least engaging posts.
Recommendations to NatGeo
- Post more wildlife, especially images featuring animals, which reliably drive the highest engagement.
- Post fewer monochrome people photos, which sit at the bottom of the engagement range.
- Write captions that support the image. Captions add predictive signal beyond the picture alone, so they shape engagement too.