Biases in recommendation systems

Previous-model bias
From Liu et al. (2017) :

"During our efforts to improve ranking, we experienced a major challenge. Because engagement logs are used for training, we introduced a direct feedback loop, as described in [22]: the model that is currently deployed dramatically impacts the training examples produced for future models. We directly observed the negative impact of this feedback loop. When we trained the first ranking model, the logs reflected user’s engagement with results ranked only by the candidate generator. The learned model was applied to rank these same candidates. Over the following months, the training data only reflected engagement with pins that were highly ranked by the existing model (Figure 6a). When we tried to train a model with the same features but with the latest engagement data, we were unable to beat the already- deployed model. We hypothesized that the feedback loop posed a problem since the distribution of training pins no longer matched the distribution of pins ranked at serving time."

Solutions
From Liu et al. (2017) :

"To alleviate this “previous-model” bias in the training data, we allocate a small percentage of traffic for “unbiased data collection”: for these requests, we show a random sample from all our candidate sources, randomly ordered without ranking. This isolates the training data from being influenced by the previous model (Figure 6b). Although the unranked results are lower quality, they provide valuable data for training new ranking models. To avoid degrading any particular user’s experience too much, each user is served unranked pins on only a small random subset of queries. Although the volume of training data becomes limited to this small percentage of overall traffic, the resulting models perform better than models trained with biased data."

Position bias
Counter-factual reasoning:
 * https://dl.acm.org/citation.cfm?id=3159732
 * source code: https://www.cs.cornell.edu/people/tj/svm_light/svm_proprank.html
 * https://www.cs.cornell.edu/people/tj/publications/agarwal_etal_18c.pdf