Insights from users through analysis of user-generated images

To create user personas, we analyse users' interactions with the image clusters, grouping similar users together based on their preferences and interests, and identifying the dominant topics of each user persona by looking at the captions associated with the image clusters. More specifically, we create user embeddings based on each user’s interactions with the image clusters. These user embeddings are then once again clustered to group similar users together to form personas. ‍

These clusters can then identify the personas by looking at how much each image cluster attributes to the average user embedding of each cluster. This way the most dominant topics of each user persona can be identified by the textual content of the captions associated with the image clusters most contributing to that user persona.

By understanding the personas of their users, Stampix can make more informed decisions about which storylines and resources to suggest to each user. For example, a user persona that is interested in travel and nature photography may be more likely to be interested in a photobook with a storyline about a nature hike or a trip to a national park. On the other hand, a user persona that is interested in family and children photography may be more likely to be interested in a photobook with a storyline about a family reunion.

These insights can then be used to generate photobooks using sets of key sentences, sets of example images and target dates describing a storyline aligned with the detected photobook personas. 

This approach allows Stampix not only to automatically generate personalised photobooks for each user without the need for a human editor, but also to understand the interests and preferences of their users on a much more foundational level. This improves the user experience, generates more value for their customers and serves as a groundwork for many future iterations.

We would like to thank Stampix for the opportunity to work with them on this project. ML6 will continue to provide expert support where needed.

ed deep learning models to extract the embeddings for each of the user’s photos.

Next, we annotate the image clusters by embedding popular image captions taken from an existing large dataset containing millions of image captions to the same embedding space as the anonymised image embeddings. We can then calculate the closest caption embedding to each of the clusters’ centres which results in a set of predicted captions for each image cluster.