These are the fascinating AI research papers that have been published this year. It mixes artificial intelligence (AI) with discoveries in data science. It is ordered chronologically and includes a link to a longer article.

Conditional Generative Model-based Predicate-Aware Query Approximation

Approximate Query Processing (AQP) aims to offer rapid yet "exact enough" responses for expensive aggregate searches, thus improving user experience in the interactive exploration of massive datasets. Compared to typical query processing on database clusters, recently developed Machine-Learning-based AQP approaches can give very low latency because query execution only includes model inference. However, as the number of filtering predicates (WHERE clauses) increases, so does the approximation inaccuracy for these approaches. Furthermore, analysts frequently use queries with many predicates to identify insights. As a result, keeping approximation error low is critical to preventing analysts from reaching incorrect conclusions.

In this work, the researchers offer ELECTRA, a predicate-aware AQP system capable of answering analytics-style queries with many predicates with substantially fewer approximation errors. ELECTRA employs a conditional generative model, which learns the conditional distribution of the data and generates a small (1000 row) but representative sample at runtime, on which the query is executed to get the estimated result. Compared to baselines, their analyses with four distinct baselines on three real-world datasets reveal that ELECTRA yields lower AQP error for many predicates.

Deep Clustering of Text Representations for Supervision-Free Probing of Syntax

The researchers investigate deep text representation clustering for unsupervised model interpretation and syntax induction. Because these representations are high-dimensional, standard approaches such as KMeans need to function more effectively. As a result, their technique converts and clusters the representations in a lower-dimensional cluster-friendly environment. In this paper, the researchers explore two types of syntax: part of speech induction (POSI) and constituency labeling (CoLab). Interestingly, they discover that Multilingual BERT (mBERT) has a surprising amount of English syntactic knowledge, potentially even as much as English BERT (EBERT). Their model can be utilized as a supervision-free probe, potentially less biased. 

The researchers discovered that unsupervised probes benefit from greater layers than supervised probes. They also point out that our unsupervised probe uses EBERT and mBERT representations differently, particularly for POSI. Finally, the researchers demonstrate the effectiveness of our probe as an unsupervised syntax induction technique. Their probe works well for both syntactic formalisms by modifying the input representations. The researchers report competitive results on 45-tag English POSI, cutting-edge performance on 12-tag POSI across ten languages, and competitive results on CoLab. The researchers also execute zero-shot syntax induction on resource-poor languages, with promising results.

FairFoody: Bringing in Fairness in Food Delivery

Along with the increasing growth and importance of food delivery platforms, concerns have arisen regarding the employment conditions of the gig workers that drive this growth. Their research of data from a real-world food delivery network in three major Indian cities reveals vast disparities in the wages of delivery agents.

The researchers formulate the challenge of fair pay distribution among agents while ensuring timely meal delivery in this work. They show that not only is the issue NP-hard, but it is also intractable in polynomial time. The researchers overcame this processing restriction by developing FairFoody, a revolutionary matching technique. Extensive trials on real-world food delivery datasets reveal that FairFoody improves equitable income distribution up to tenfold compared to baseline techniques while not influencing the customer experience.