These are the year’s most intriguing AI research articles to be published. It combines AI and data science breakthroughs. It is chronologically organised and contains a link to a longer article.
The increasing availability of cell phones has made it possible for non-profits to convey vital health information to their beneficiaries promptly. This paper describes their efforts to support non-profit organisations that use automated message programmes to give preventative care information to beneficiaries (new and expectant moms) during pregnancy and after delivery. Unfortunately, many programme participants leave, posing a significant obstacle to such information distribution operations. Yet, non-profits frequently have limited health-worker resources (time) to place vital service calls for live interaction with beneficiaries to avert engagement declines.
The researchers devised the Restless Multi-Armed Bandits (RMABs) system to aid non-profits in optimising this restricted resource. The innovative clustering of offline historical data to infer unknown RMAB parameters is a significant technical contribution of this system. Their second key contribution is evaluating their RMAB system in partnership with a non-governmental organisation by studying real-world service quality enhancement. Finally, the study examined tactics for optimising service calls to 2,303,3 participants over seven weeks to reduce engagement drops.
The researchers demonstrate that the RMAB group significantly outperforms other comparison groups, lowering engagement declines by 30%. According to the authors’ understanding, this is the first study to demonstrate the efficacy of RMABs in actual public health situations. The researchers are transferring their RMAB technology to the NGO for application in the real world.
How Private Is Your RL Policy? An Inverse RL-Based Analysis Framework
Reinforcement Learning (RL) enables agents to learn how to accomplish various tasks from scratch. In fields such as autonomous driving and recommendation systems, optimal RL policies that memorise any portion of the private reward could result in a privacy breach. Existing differentially-private RL policies produced from RL algorithms, such as Value Iteration, Deep Q Networks, and Vanilla Proximal Policy Optimization, are investigated.
The researchers offer a new Privacy-Aware Inverse RL (PRIL) analytical framework, which implements reward reconstruction as an adversarial attack on private policies that agents may implement. For this purpose, they offer the reward reconstruction attack, in which we use an Inverse RL algorithm to attempt to rebuild the original reward from a privacy-preserving policy. If the agent employs a stringently private policy, the opponent will have difficulty reconstructing the original reward function. Using this approach, the researchers experimentally evaluate the efficacy of the privacy guarantee provided by the private algorithms on several instances of the FrozenLake domain with differing degrees of complexity. Based on their investigation, the researchers conclude that the current level of privacy granted needs to be revised to preserve reward functions in RL. The researchers do this by quantifying the degree to which each private policy safeguards the reward function by calculating the distance between the original and rebuilt rewards.
Interpretable Neural Subgraph Matching for Graph Retrieval
A graph retrieval system seeks to deliver the most pertinent corpus graphs given a query graph and a database of corpus graphs. Numerous industries use graph retrieval based on subgraph matching, including molecular fingerprint detection, circuit design, software analysis, and question-answering. A corpus graph is relevant to a query graph in these applications if the query graph is (precisely or roughly) a subgraph of the corpus graph. Existing neural graph retrieval models compare the node or graph embeddings of the query-corpus pairings to calculate their relevance scores. However, the query and corpus graphs might need edge consistency with such models. They also frequently employ symmetric relevance scores, which are inappropriate for subgraph matching because the underlying relevance score for subgraph search should be calculated using the partial order imposed by the subgraph-supergraph link. They consequently exhibit poor retrieval performance when subgraph matching is involved.
To combat this, the researchers suggest ISONET, a brand-new interpretable neural edge alignment formulation that can better learn the edge-consistent mapping required for subgraph matching. Concerning subgraph matching, ISONET now uses a new scoring system that enforces an uneven relevance score. However, because of the way ISONET is built, it can quickly determine which subgraph in a corpus graph pertains to a specific query graph. On various datasets, their investigations demonstrate that ISONET outperforms current graph retrieval formulations and techniques. Despite being trained simply on binary relevance labels of entire graphs and lacking fine-grained ground truth data on the node or edge alignments, ISONET may still offer interpretable alignments between query-corpus graph pairings during inference.
