product analyticsimage recognitionsales datacomputer visiondata sciencemachine learning
Creating Multimodal Product Analytics: Combining Image Recognition with Sales Data to Identify Visual Patterns that Drive Revenue
7 min read · 1,297 wordsBy Orandi Felix
The script assumes `data/images/` contains all product photos and `data/sales.csv` has SKU columns. ResNet18 is lightweight (200MB) and works on CPU — swap to ResNet50 or EfficientNet-B3 if you need higher fidelity and have a GPU.
Always sanity-check embeddings. Plot the UMAP clusters and overlay sample images (use `matplotlib.pyplot.imshow`). If visually dissimilar products cluster together, the model is picking up camera artifacts, not product features.
Before starting, ask: "Could a human look at two products and guess which sells more?" If the answer is "only if they know price/discount," image features won’t help.