skip to main content

Title: Estimating Error and Bias in Offline Evaluation Results
Offline evaluation protocols for recommender systems are intended to estimate users' satisfaction with recommendations using static data from prior user interactions. These evaluations allow researchers and production developers to carry out first-pass estimates of the likely performance of a new system and weed out bad ideas before presenting them to users. However, offline evaluations cannot accurately assess novel, relevant recommendations, because the most novel recommendations items that were previously unknown to the user; such items are missing from the historical data, so they cannot be judged as relevant. A breakthrough that reliably produces novel, relevant recommendations would score poorly with current offline evaluation techniques. While the existence of this problem is noted in the literature, its extent is not well-understood. We present a simulation study to estimate the error that such missing data causes in commonly-used evaluation metrics in order to assess its prevalence and impact. We find that missing data in the rating or observation process causes the evaluation protocol to systematically mis-estimate metric values, and in some cases erroneously determine that a popularity-based recommender outperforms even a perfect personalized recommender. Substantial breakthroughs in recommendation quality, therefore, will be difficult to assess with existing offline techniques.
Authors:
;
Award ID(s):
1751278
Publication Date:
NSF-PAR ID:
10146883
Journal Name:
Proceedings of the 2020 Conference on Computer-Human Interaction and Information Retrieval
Page Range or eLocation-ID:
392 - 396
Sponsoring Org:
National Science Foundation
More Like this
  1. Today’s recommender systems are criticized for recommending items that are too obvious to arouse users’ interest. That is why the recommender systems research community has advocated some ”beyond accuracy” evaluation metrics such as novelty, diversity, coverage, and serendipity with the hope of promoting information discovery and sustain users’ interest over a long period of time. While bringing in new perspectives, most of these evaluation metrics have not considered individual users’ difference: an open-minded user may favor highly novel or diversified recommendations whereas a conservative user’s appetite for novelty or diversity may not be that large. In this paper, we developedmore »a model to approximate an individual’s curiosity distribution over different levels of stimuli guided by the well-known Wundt curve in Psychology. We measured an item’s surprise level to assess the stimulation level and whether it is in the range of the user’s appetite for stimulus. We then proposed a recommendation system framework that considers both user preference and appetite for stimulus where the curiosity is maximally aroused. Our framework differs from a typical recommender system in that it leverages human’s curiosity to promote intrinsic interest with the system. A series of evaluation experiments have been conducted to show that our framework is able to rank higher the items with not only high ratings but also high response likelihood. The recommendation list generated by our algorithm has higher potential of inspiring user curiosity compared to traditional approaches. The personalization factor for assessing the stimulus (surprise) strength further helps the recommender achieve smaller (better) inter-user similarity.« less
  2. Traditional offline evaluations of recommender systems apply metrics from machine learning and information retrieval in settings where their underlying assumptions no longer hold. This results in significant error and bias in measures of top-N recommendation performance, such as precision, recall, and nDCG. Several of the specific causes of these errors, including popularity bias and misclassified decoy items, are well-explored in the existing literature. In this paper we survey a range of work on identifying and addressing these problems, and report on our work in progress to simulate the recommender data generation and evaluation processes to quantify the extent of evaluationmore »metric errors and assess their sensitivity to various assumptions.« less
  3. Recent work in recommender systems has emphasized the importance of fairness, with a particular interest in bias and transparency, in addition to predictive accuracy. In this paper, we focus on the state of the art pairwise ranking model, Bayesian Personalized Ranking (BPR), which has previously been found to outperform pointwise models in predictive accuracy, while also being able to handle implicit feedback. Specifically, we address two limitations of BPR: (1) BPR is a black box model that does not explain its outputs, thus limiting the user's trust in the recommendations, and the analyst's ability to scrutinize a model's outputs; andmore »(2) BPR is vulnerable to exposure bias due to the data being Missing Not At Random (MNAR). This exposure bias usually translates into an unfairness against the least popular items because they risk being under-exposed by the recommender system. In this work, we first propose a novel explainable loss function and a corresponding Matrix Factorization-based model called Explainable Bayesian Personalized Ranking (EBPR) that generates recommendations along with item-based explanations. Then, we theoretically quantify additional exposure bias resulting from the explainability, and use it as a basis to propose an unbiased estimator for the ideal EBPR loss. The result is a ranking model that aptly captures both debiased and explainable user preferences. Finally, we perform an empirical study on three real-world datasets that demonstrate the advantages of our proposed models.« less
  4. Collaborative filtering algorithms find useful patterns in rating and consumption data and exploit these patterns to guide users to good items. Many of these patterns reflect important real-world phenomena driving interactions between the various users and items; other patterns may be irrelevant or reflect undesired discrimination, such as discrimination in publishing or purchasing against authors who are women or ethnic minorities. In this work, we examine the response of collaborative filtering recommender algorithms to the distribution of their input data with respect to one dimension of social concern, namely content creator gender. Using publicly available book ratings data, we measuremore »the distribution of the genders of the authors of books in user rating profiles and recommendation lists produced from this data. We find that common collaborative filtering algorithms tend to propagate at least some of each user’s tendency to rate or read male or female authors into their resulting recommendations, although they differ in both the strength of this propagation and the variance in the gender balance of the recommendation lists they produce. The data, experimental design, and statistical methods are designed to be reusable for studying potentially discriminatory social dimensions of recommendations in other domains and settings as well.« less
  5. Recommender systems learn from past user preferences in order to predict future user interests and provide users with personalized suggestions. Previous research has demonstrated that biases in user profiles in the aggregate can influence the recommendations to users who do not share the majority preference. One consequence of this bias propagation effect is miscalibration, a mismatch between the types or categories of items that a user prefers and the items provided in recommendations. In this paper, we conduct a systematic analysis aimed at identifying key characteristics in user profiles that might lead to miscalibrated recommendations. We consider several categories ofmore »profile characteristics, including similarity to the average user, propensity towards popularity, profile diversity, and preference intensity. We develop predictive models of miscalibration and use these models to identify the most important features correlated with miscalibration, given different algorithms and dataset characteristics. Our analysis is intended to help system designers predict miscalibration effects and to develop recommendation algorithms with improved calibration properties.« less