Community Blog "Ask Me Anything" with Hongxia Yang, a Judge of KDD Cup 2020 Challenges from Alibaba DAMO Academy

"Ask Me Anything" with Hongxia Yang, a Judge of KDD Cup 2020 Challenges from Alibaba DAMO Academy

In this blog, we interviewed Dr. Hongxia Yang, a senior staff data scientist at Alibaba DAMO Academy, about the recent KDD Cup 2020 and the trends in debiasing.

KDD Cup 2020 Challenges for Modern E-Commerce Platform, sponsored by Alibaba, Alibaba DAMO Academy, Duke University, Tsinghua University and UIUC and hosted by Alibaba Tianchi Platform, have announced the final result (See top 20 lists: Debiasing Track, Multimodalities Recall Track). The two tasks of the competition, recall for multi-modal entities and debiasing, are designed to help tackle some fundamental challenges and to further nourish the development for the search and recommender systems of e-commerce and retail companies.

During the competition, Dr. Hongxia Yang, a judge of the competition from Alibaba DAMO Academy was invited to have an "Ask Me Anything" session to answer some questions from participants and give some insights about the background and value of the competition.

Dr. Hongxia Yang is a Senior Staff Data Scientist of Data Analytics and Intelligence Lab at Alibaba DAMO Academy. She received her PhD in Statistics from Duke University in 2010. Her interests span the areas of Bayesian statistics, time series analysis, spatial-temporal modeling, survival analysis, machine learning, data mining (and their applications to problems in business analytics), and big data. She used to work as the Principal Data Scientist at Yahoo! Inc. and Research Staff Member at IBM T.J. Watson Research Center. She has published over 40 top conference and journal papers and is serving as the associate editor for Applied Stochastic Models in Business and Industry. She has been selected as an Elected Member of the International Statistical Institute (ISI) in 2017.

Below are selected Q&A from the AMA with Dr. Yang. You can also view the full version of the AMA by clicking the links below:

Q: What is your main research focus at DAMO Academy?

A: We are focusing on the cognitive intelligence that is applicable to the recommender systems. Cognitive intelligence mainly includes three parts, cross-domain knowledge graph, graph neural network (GNN) based inductive reasoning platform, and user-interacted content (nlp/videl/image) understanding; with their applications to the modern recommender systems.

Q: The track of Debiasing is focusing on the fairness of exposure. How does exposure bias happen on a recommender system?

A: It is more related to the nature of the big data algorithm, which pays more attention to those regions that are abundant with data so the algorithms are more confident to issue the results. There is currently a very popular trending in machine learning to solve this problem, fairness in ML.

Q: What problem will exposure bias cause to e-commerce platforms?

A: Exploration for long-tailed/cold-start items will be not sufficient.

Q: Why is NDCG@50 used to measure the result of the competition? Is there any other method being used to measure the fairness of exposure in practice?

A: Indeed, ndcg@50 is an offline measure and recommender systems are an online changing environment. In practice, we will also pay more attention to the total number of items that are exposed in the long term, for example, one month.

Q: Does DAMO Academy have any progress on reducing the exposure bias during the past few years?

A: Yes, we have one paper in submission for the NeurIPS, through combining a contrastive learning framework with the recommender system to explore more efficiently for those long-tailed items.

Q: Will the short-term goals like ctr, cvr, or gmv (as mentioned in the competition introduction) be affected when reducing the exposure bias with algorithms? How can we balance them?

A: Yes, indeed. We also consider long-term goals, e.g., the total number of items that are exposed in a long period (one month for example) and the distribution of exposed items in different categories. We should also pay attention to the users' duration time.

Q: What's your expectation on the participants' results through this challenge?

A: Current modern recommender systems are far more complex and attractive compared to the traditional ones. We need to recommend not only items, but videos, texts, and topics, among several other formats. We would like to nourish the community to access the ongoing challenges and make progress in the interdisciplinary, which we believe will have breakthroughs both in academia and the industry.

0 0 0
Share on

Alibaba Clouder

2,600 posts | 750 followers

You may also like


Alibaba Clouder

2,600 posts | 750 followers

Related Products