Report on AAAI‑2017 Conference Highlights and Ctrip’s Hybrid Collaborative Filtering Model

Author Biography – Wu Zhonghuo is a senior algorithm engineer at Ctrip’s Technology Center, focusing on personalized recommendation and natural language processing since joining in 2013.

Conference Overview – AAAI‑2017 was held in San Francisco from February 4‑9, 2017, during the rainy season. The venue featured the iconic Golden Gate Bridge shrouded in fog. AAAI, founded in 1979, is a premier AI conference with over 4,000 members and 230+ fellows.

Program Highlights – The tightly packed schedule included tutorials, invited talks, panels, AI‑in‑Practice sessions, technical talks, poster/demo sessions, workshops, and a doctoral consortium. Notable AI‑in‑Practice speakers were senior leaders from Google, Facebook, Amazon, and other tech giants.

Selected Talks – Google Brain’s Vincent Vanhoucke discussed image, speech, and machine‑translation research and robot perception. Amazon’s Alex Smola presented “Scalable and Personal Deep Learning with MXNet,” comparing MXNet with Caffe and TensorFlow. Facebook’s Joaquin Quinonero Candela described large‑scale AI systems for vision, text, and speech, showcasing the Lumos platform.

Keynote & Panels – MIT Media Lab’s Rosalind Picard delivered a keynote on emotion‑intelligence technologies, demonstrating facial‑expression recognition, smartwatch stress monitoring, and AI‑based depression prevention. Panels covered AI ethics, AI for education, AI for social good, expert systems history, and AI in poker.

Award‑Winning Papers – The best paper award went to Stanford’s Russell Stewart and Stefano Ermon for “Label‑Free Supervision of Neural Networks with Physics and Domain Knowledge,” which trains CNNs without explicit labels by enforcing physical constraints.

Ctrip’s Contribution – Ctrip presented a paper titled “A Hybrid Collaborative Filtering Model with Deep Structure for Recommender Systems.” The model introduces an Additional Stacked Denoising Auto‑Encoder (aSDAE) that learns user and item latent vectors from rating lists, incorporating side‑information similar to sequence‑to‑sequence models.

The aSDAE outputs are combined with matrix factorization; two aSDAEs learn user and item embeddings, whose inner product approximates the original rating matrix. The joint objective (matrix‑factorization loss + two aSDAE losses) is optimized via SGD, yielding superior performance on multiple datasets.

Impact – The hybrid model addresses severe data sparsity in OTA (online travel agency) scenarios, has been deployed in over 50 personalized recommendation contexts, and in some cases increased conversion rates by up to 13 times, significantly enhancing user travel experience.

Future Directions – Beyond recommendation, Ctrip plans to deepen research in natural language processing, computer vision, and speech, aiming to showcase more AI breakthroughs at top conferences.

Call for Interest – Readers interested in the Ctrip paper can request the full manuscript via the Ctrip Technology Center WeChat public account.