Notebooks and examples on how to onboard and use various features of Amazon Personalize
Open the getting_started folder to find a CloudFormation template that will deploy all the resources you need to build your first campaign with Amazon Personalize. The notebooks provided can also serve as a template to building your own models with your own data.
This repository is cloned into the environment so you can explore the more advanced notebooks with this approach as well.
If you just want a simple walkthrough to explore later you can execute personalize_sample_notebook.ipynb, it works well inside the same Jupyter environments.
Collaborative filtering based on user-item interaction tables. The intuition behind is that similar users like similar items.
Hybrid recommendation also considering user, item, and event meta-data. The result is to extrapolate to out-of-sample users and items, based on their meta-data features.
You have some historical data and you want to know how personalize performs on your data. Here is what we suggest:
- Temporally split your data into a 'past' training set and a 'future' testing set.
- Upload the 'past' data to Amazon Personalize, train a solution, and deploy a campaign.
- Use your campaign to get recommendation for all of your users, and compare with the 'future' testing set.
This is an example to complete the steps above. We include a basic popularity-based recommendation, which should be easy to beat. This is for sanity checking purposes. A common next-step is to kepp the same training and testing splits, but train different models for more serious offline comparisons.
The 'sims' recipe allows next-item recommendation based on a single previous item. It is faster to train and easier to interpret, e.g., in the form of "you see these recommendations because you watched A". (On the contrary, 'hrnn' considers the entire user consumption histories as the recommendation contexts and can therefore be more personalized).
Similar to the 'hrnn' example, this example uploads the 'past' data from temporal splitting and evaluates the recommendation against the held-out 'future' ground truth. The results compare favorably with a popularity-based recommendation baseline. We also include examples showing that different "cause" items would lead to different 'sims' results.
Real-time personalization should respond to new click events by the user. For 'hrnn' sequence model, this is straightforward. After you 'put_events' to our system, the user states get updated and the corresponding recommendations change.
Here is an example showing how User A's recommendation will eventually look like User B's recommendation, if User B's events are appended after User A.
Meta-data is ubiquitous. User zipcodes and device types can be useful indicators of preference; item categories and tags can be useful patterns in decision making; click and purchase events may imply different utilities to the user.
This example shows how these useful information can be uploaded to our system to aid recommendation. A caveat is that the improvements of meta-data recipes depend on how much information can be extracted from the provided meta-data. Movie genres may be less useful compared with movie ratings, or better, directors and stars.
An important functionality that meta-data, particularly item meta-data, provides is to generalize to new 'cold-start' items. Examples include new releases, new products, or live items. Without personalization, a global policy to introduce these new items may incur large promotional costs. Personalized 'cold-start' helps reduce these costs.
This example shows how we may personalize item 'cold-start' by exploring only in the same movie genres that the user would be interested in. The steps are:
- Randomly hold out 50% of all items to simulate an item 'cold-start' scenario.
- Remove these items from the interactions table.
- Use temporal splitting, train a solution, and deploy a campaign with the remaining training data.
- Compute metrics on the held out items in the testing data split; these items never show up in the training split.
We can see that the cold-start recipe indeed recommends new movies in the same genres that the user prefers. As a baseline and without personalization, new movies would have a lower click rate, which implies larger promotional costs.
This sample code is made available under a modified MIT license. See the LICENSE file.