How to Build Product Recommendations With Item-Based Filtering

When should ML engineers use item-based collaborative filtering?

Use item-based collaborative filtering whenever the recommendation trigger is a known item — a product or a location — rather than a user ID. This is the single most important decision in the tourism track. A shopper viewing one product, or a tourist standing at one place, gives you an item as input, so you compute similar items from the ratings matrix (users × items) and return a ranked list. Reserve user-based filtering for cases where a logged-in user ID drives the recommendation. Defaulting to user-based when your trigger is an item is a classic mismatch that produces irrelevant recommendations.

How do you build the image classification half with transfer learning?

Never build the CNN from scratch. Organise images into one subfolder per class and load them with `image_dataset_from_directory`, which auto-assigns labels and splits training/validation. Load VGG16 with include_top=False to strip the ImageNet head, flatten the output, add Dense + ReLU, add a Dropout layer (required and it curbs overfitting), then a Dense + Softmax layer with neurons equal to your number of classes. Compile with Adam, categorical_crossentropy, accuracy, and EarlyStopping. Mismatching the final layer's neuron count to your class count is the most common shape-error at training time.

Why train the classifier twice?

Because the augmentation A/B test is a required, insight-generating deliverable. Train once without augmentation layers, then again with a Sequential augmentation block — RandomFlip, RandomRotation, RandomZoom — prepended to the same architecture. Report training vs validation accuracy for both runs side by side and explain whether augmentation reduced overfitting and improved generalisation. This gives you concrete evidence about how synthetic image variation affects your model rather than a vague claim that augmentation 'usually helps'.

How do you assemble the recommendation data?

Merge the ratings CSV with the items CSV on `place_id` (or `product_id`). Run `value_counts` on the item key to find the most-visited locations or most-purchased products — this is your popularity baseline. Filter by category or search a description column for substrings (for example 'nature') to answer domain questions, and compute average ratings by city with `groupby + mean`. Then build the item-based model on the ratings matrix, so that given one item, you return the most similar items ranked by the collaborative signal across all users.

What should ML engineers watch out for in production?

Three things. First, always connect to a GPU runtime — VGG16 transfer learning is impractically slow on CPU. Second, keep the classification model and the recommender as separate parts with separate evaluation; the vision model reports accuracy, the recommender is validated on relevance. Third, treat your ratings data hygiene seriously: remove duplicates, drop ID-only columns, and handle missing ratings by dropping or flagging rather than filling with zero, since a zero rating is a real signal you don't want to invent. Getting the filtering type and the data hygiene right matters more than exotic model architecture.

Next step: Confirm your recommendation trigger — item or user — before writing any code, then load your per-class image folders and run the with/without-augmentation comparison. Choose the filtering type first; it determines your entire recommender design.