Have you ever noticed how your behavior changes based on where a product is positioned on a webpage? Well, that’s the impact of position bias on user behavior in product recommendations. When the position changes, the context changes too, leading to variations in user behavior. For example, a product’s conversion rate can vary throughout the week based on its position on a webpage.

Imagine recommending the same product in three different playlists on the same date, to the same user, but at various positions on the webpage. Despite the similar context, biases are present due to the vertical distribution of the playlists on the webpage.

Even with similar context, biases can still exist due to the way playlists are displayed. This bias needs to be acknowledged and countered in order to provide fair recommendations.

#3 Bias-aware Modelling Frameworks

To address position bias in modeling user behavior, bias-aware frameworks have been developed. These frameworks aim to not only identify bias but also incorporate it into the predictive models.

One common approach is to use the position as a feature in the predictive model. By including the product’s position as an ordinal feature during training, the model can account for position bias. However, setting the position feature to a constant during predictions ensures that all items are equally considered.

However, this approach has its limitations, as the model output can vary greatly depending on the chosen constant, leading to unstable rankings past the first position. This instability can impact user experience and click-through rates.

To address this issue, a new approach called Position-bias Aware Learning framework (PAL) was developed. PAL optimizes for position bias and interaction simultaneously, resulting in more accurate predictions.