Training and testing of recommender systems on data missing not at random

By Harald Steck. 2010

425 citations recorded by the source corpus.

Subfields: Artificial Intelligence, Information Systems, Management Science and Operations Research

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Information Systems

In Management Science and Operations Research

Citations in the corpus

Gaussian Ranking by Matrix Factorization (2015, methodology): “There has been a large body of work on training collaborative ﬁltering approaches as to maximize various ranking metrics, see eg [5, 8, 17, 2, 18, 10, 20, 21, 14, 13, 12].”
Recommendations as Treatments: Debiasing Learning and Evaluation (2016, background): “1 ratings Figure shows the matrix of true Y ∈ U × I for our toy example, where a sub-5”
The Selective Labels Problem: Evaluating Algorithmic Predictions in the Presence of Unobservables (2017, background): “(3) missing not at random (MNAR) [40], where the variable value that is missing is related to the reason why it’s missing.”
Modeling Sentences in the Latent Space (2012, methodology): “In the RS setting, they are interested in predicting individual ratings, while we are interested in the sentence 1 An efﬁcient way to compute equation 4 is proposed in (Steck, 2010). semantics.”
Bias and Debias in Recommender System: A Survey and Future Directions (2020, methodology): “[21], [63] directly impute the missing data with a specific value r0 and optimize the model with the following objective function:”
Transparent, Scrutable and Explainable User Models for Personalized Recommendation (2019, background): “In practice, these are biased towards items users like or know [29, 45].”
Recommender Systems (2016, positive): “This property of rating distributions is also known as Missing Not At Random (MNAR), or selection bias [402, 565].”
Calibrated recommendations (2018, methodology): “…we learned a weighted 50-dimensional matrix-factorization model on these training data following the usual approaches described in [14, 18, 23], where we tuned the hyper-parameters, i.e., the L2-norm regularization and the weight for the missing plays (negative sampling), as to maximize…”
Evaluation of recommendations: rating-prediction and ranking (2013, methodology): “Due to this diﬀer-ence, in [22] AllRank was found to achieve a considerably larger top-N hit-rate or recall when ranking all items in the collection, compared to various state-of-the-art approaches optimized on the observed ratings only, see results in [22] and compare to [10, 4].”
Multivariate Time Series Anomaly Detection and Interpretation using Hierarchical Inter-Metric and Temporal Embedding (2021, methodology): “, recall) for recommendations [29], we propose a new metric, namely “InterPretation Score” (IPS), to evaluate the anomaly interpretation accuracy at the segment level.”
Probabilistic Matrix Factorization with Non-random Missing Data (2014, methodology): “Steck (2010) does not learn a generative model for the data and optimizes a metric that is robust to MNAR data.”
Top-N Recommendation with Missing Implicit Feedback (2015, methodology): “To address the MNAR problem in explicit feedback systems, a missing data model was proposed in [8], and it was shown that the Top-N recall and the Area-under-the-TopN-curve(ATOP) measures evaluated on the observed data provided an unbiased estimate of performance on the complete data under the missi”
Human Interaction with Recommendation Systems: On Bias and Exploration (2017, background): “ treat the data as xed, rather than dynamic. There has been less work on selection bias, which was rst demonstrated by Marlin [2003], and subsequent work [Amatriain et al., 2009, Marlin et al., 2007, Steck, 2010]. Rather than modeling user behavior directly, they impose the statistical assumption of”
A survey of collaborative filtering based social recommender systems (2014, background): “For instance, the rating values range from 1 to 5 stars in the Netflix [19] data, and the author in [9] considered 5-star ratings as relevant (i.”
Ranking with non-random missing ratings: influence of popularity and positivity on evaluation metrics (2012, methodology): “In closer relation to our work, [11] gave the following argument in favor of considering missing as negatives: if a measure ignores missing ratings, it does not penalize a system which ranks unrated items first, even though we may have the prior knowledge that people tend to rate items they like.”
Fast and Accurate Matrix Completion via Truncated Nuclear Norm Regularization (2013, background): “This problem occurs in many real applications in computer vision and pattern recognition, such as image inpainting [8], [9], video denoising [10], and recommender systems [11], [12].”
Collaborative Filtering-Based Recommendation of Online Social Voting (2017, methodology): “In Section IV-B, we propose a multichannel MF model, which factorizes user-voting inter- actions, user–user interactions, and user-group interactions simultaneously, gearing to optimize top-k hit rate.”
Collaborative Denoising Auto-Encoders for Top-N Recommender Systems (2016, background): “These models are not suitable for top-N recommenders [13, 23].”