Insights from AAAI 2018: Conference Overview, Paper Highlights, and Ant Financial Contributions

AAAI (Association for the Advancement of Artificial Intelligence) is one of the premier academic organizations in AI, and its annual conference (AAAI) is a top‑tier global AI event.

Founded in 1979, AAAI has over 6,000 members and has been chaired by AI pioneers such as Allen Newell, Edward Feigenbaum, Marvin Minsky, and John McCarthy.

In 2018 the conference received 3,808 paper submissions and accepted 938 papers, a 47% increase over the previous year.

China submitted the most papers, ranking first worldwide, while the United States had slightly more accepted papers.

Machine learning methods attracted the highest number of submissions and acceptances, with the image sub‑field showing a 257% increase in submissions and a 285% increase in acceptances compared to the prior year.

Tracks with higher acceptance rates included computational sustainability, reasoning under uncertainty, and cognitive modeling, whereas AI applications and machine learning applications had lower acceptance rates.

The best paper award went to "Memory‑Augmented Monte Carlo Tree Search" from the University of Alberta (link: https://webdocs.cs.ualberta.ca/~mmueller/ps/2018/Chenjun-Xiao-M-MCTS-aaai18-final.pdf). The best student paper was "Counterfactual Multi‑Agent Policy Gradients" from Oxford University (link: https://arxiv.org/abs/1705.08926). The classic paper award recognized the influential 2000 AAAI paper "PROMPT: Algorithm and Tool for Automated Ontology Merging and Alignment".

Paper Highlights:

• GraphGAN: Graph Representation Learning with Generative Adversarial Nets – introduces a generator G(V|Vc) and discriminator D(V|Vc) trained via a min‑max game, proposes Graph Softmax to overcome softmax limitations, and demonstrates superior performance on five public datasets.

• HARP: Hierarchical Representation Learning for Networks – proposes hierarchical graph coarsening (edge collapse and star collapse) to learn node embeddings more efficiently; experiments on DBLP, Blogcatalog, and CiteSeer show consistent improvements.

• PrivSR: Privacy‑Preserving Social Recommendation – a semi‑decentralized recommendation framework that protects sensitive user data using differential privacy; experiments on Ciao and Epinions datasets show superior performance as more data become private.

• Deep Asymmetric Transfer Network for Unbalanced Domain Adaptation – combines unsupervised auto‑encoder loss, supervised soft‑max loss, and a linear transformation G to align source and target domains; optimized via block coordinate descent and validated on two public datasets.

Invited presentations included a talk by Zoubin Ghahramani on probabilistic machine learning and AI, covering deep learning history, limitations, Bayesian deep learning, and Uber’s Pyro framework.

Ant Financial Contributions:

• cw2vec: Learning Chinese Word Embeddings with Stroken‑grams – a new algorithm leveraging Chinese character stroke information, outperforming word2vec, GloVe, and CWE on public benchmarks.

• Privacy Preserving Point‑of‑Interest Recommendation Using Decentralized Matrix Factorization – proposes a decentralized matrix factorization method that keeps user data on personal devices, reducing storage/computation costs and protecting privacy.

In conclusion, the Ant Financial team gained valuable insights and cutting‑edge algorithms from AAAI 2018, and looks forward to applying these ideas to real‑world scenarios within Alibaba and Ant Financial, while hoping for more Chinese innovations at future top AI conferences.