HeatWave Release Notes
The following recommendation model types are available:
Recommendation models that use explicit feedback collect data on users that directly provide ratings on items. The user ratings can be positive or negative. The recommendation models then use the feedback to generate predicted ratings for users and items. The ratings are specific values, and the higher the value, the better the rating. See Recommendation Models to review models that support explicit feedback.
Recommendation models that use implicit feedback collect data on users' behavior, such as past purchases, clicks, and view times. Users do not have to explicitly express their taste about an item. When a user interacts with an item, the implication is that they prefer it to an item that they do not interact with. Therefore, only positive observations are available. The non-observed user-item interactions are a blend of negative feedback (the user doesn't like the item) or missing values (the user might be interested in the item). The recommendation model generates rankings for users and items. Rankings are a comparative measure, and the lower the value, the better the ranking. Because A is better than B, the ranking for A has a lower value than the ranking for B. HeatWave AutoML derives rankings based on ratings from implicit feedback for all ratings that are at or above the feedback threshold.
Implicit feedback data can be in the following formats:
Unary data: Only records if an interaction occurred or not. This type of data often uses a value of 1 to represent an interaction, such as a click or view. Non-interactions can be represented by a value of 0 or missing values.
Binary data: Explicitly categorizes interactions as positive or negative, such as users expressing likes or dislikes.
Numerical data: Provides more granular information about the interaction, such as how long a user watched a video or how many times a user listened to a song. If numerical data is used for implicit feedback, it is important to set the
feedback_thresholdoption during training to distinguish what constitutes positive feedback. This threshold determines what value is equivalent to a positive interaction. For example, if users are tracked by how many times they have interacted with an item, you might set thefeedback_thresholdwith a value of 3, which means that positive feedback is represented by users that interact with the item more than three times.
Implicit feedback uses BPR: Bayesian Personalized Ranking from Implicit Feedback, which is a matrix factorization model that ranks user-item pairs. The recommendation models use this data to generate rankings for users and items.
Content-based recommendation models allow you to include item descriptions in the input of the recommendation model. This helps the model provide more accurate representations of items. Currently, content-based recommendation models can only be used with implicit feedback. When training a content-based recommendation model, you can use the Collaborative Topic Regression model, which combines the ideas of matrix factorization models and topic modeling using Latent Dirichlet Allocation (LDA).