What it does
This KPI uses a deep learning model to predict the likelihood of an image or video frame engaging viewers on social media, measured through interactions like likes, comments, and shares.
Why it matters
High engagement potential of assets is crucial for stimulating viewer interaction, sparking curiosity, and encouraging sharing. This not only boosts brand exposure but also enhances overall reach and can lead to more marketing successes.
How it works
The deep learning model assesses the engagement potential of an asset, scoring from 0 (low potential) to 100 (high potential). This score is contextualized against norms for different asset types to classify it as low, medium, or high engagement potential.
How to achieve great results
The illustration shows sample images with model predictions. Top left image has lowest, bottom right image has highest engagement potential.
Images with high activation potential feature vibrant colors, unique perspectives, and human presence or activities that evoke emotion or tell a story. In contrast, low-engagement images often have dull colors, lack clear focal points, or depict mundane subjects without emotional or narrative appeal.
Examples with assets:
Background on AI model training
Human ground truth engagement data to train image popularity models are commonly scraped from social media platforms like Instagram or Flickr. There are a few caveats however on how to work with this kind of data.
- One is that the number of followers of an account will clearly bias the number of likes or comments any given image will generate. The more followers, the more likely an image will get comparably more likes.
- Another caveat is that even when controlling for level of follower amounts, we also need to control for the time lag between two image uploads, since if two images are one year apart they may get different likes simply because the account has more followers after one year.
The scientific best practice is therefore to control for number of followers and time between uploads and split images into pairs (“low” vs “high” popularity). To achieve this, datasets are split so that
- Each group of images belongs to the same account
- Post timestamp of any pairs of images is within one week to avoid time-changing factors such as number of followers
- Group images into “low” vs “high” popularity through the constraint:
c = |l_a – l_b| / max(l_a,l_b) >= 0.5
This ensures that the less popular image has at most half the number of likes as the more popular one. In other words, the popularity of one image must be at least 50% greater than the other to satisfy this constraint
Interestingly, when such pairs of low vs high popular images are presented to human respondents with the task to select the more popular one, performance is far from perfect and underperforms current state-of-the-art deep learning models. For example, performance of human raters was just 59% correct (chance level: 50%) in one study (Ji et al. 2023) and 72% in another (Ding et al., 2019). In contrast, current image popularity models (including aimpowers’) have accuracies of 77% and higher, thus surpassing human raters.
AI model performance
Engagement potential models at aimpower are based on the latest deep learning frameworks used in scientific literature for the purpose of emotion modelling, including convolutional deep learning networks and vision transformers. The images and engagement data used to train the models originate from Instagram and Flickr. In total we leverage 1m+ images for pretraining and 100K+ images for fine-tuning our models on the engagement prediction task. In the evaluation of the predictive model's performance, a t-test of independence was conducted to assess the relationship between the binned categories of the model's predictions and the actual values of the dependent variable. The t-test yielded a very strong statistical association between model predictions and the actual popularity scores (p<0.00001), suggesting that the model's predictions are significantly related to the observed outcomes. Our model achieves 81% accuracy overall, achieving the same level as current state-of-the-art scientific deep learning models (Ji et al. 2023) and significantly outperforming human raters (72% accuracy).