Stage 1: Monitoring & Handling Outliers Interactions Data
Stage 2: Analysing and Monitoring Position Bias
Position bias is a phenomenon commonly observed in recommender systems where items displayed at the top of a list (e.g., search results, recommendation list) are more likely to be clicked on or selected than items displayed lower down the list, even if the lower ranked items might be a better match for the user’s preferences.
This bias occurs because users tend to pay more attention to items displayed at the top of the list and are more likely to interact with those items. As a result, the top-ranked items may receive a disproportionate amount of user feedback (whether the item was relevant or not), which can in turn reinforce their ranking, perpetuating the bias.
Position bias can lead to suboptimal recommendations and lower overall system performance. Below we will look at different aspects of positional bias.
Stage 3: Optimising a solution for an additional objective
The primary objective of Amazon Personalize is to predict the most relevant polls for our users based on historical and real-time interactions data. These are the items our users will most likely interact with (for example, the polls they will most likely hunch). If we have an additional objective, such as maximising polls from Hunch App Influencers, we can create a solution that generates recommendations based on both relevance and our custom objective.
Stage 4: Fine-tune Featurization Hyperparameters