[Adapt] [Seminar] Counterfactual reasoning: introduction and application in NLU/recommendation

[龚珊三]

Hi Adapters,

It's easy for a deep learning model to learn the association of input X and output Y, but the model might miss the causality between X and Y. 
For example, in school, myopic classmates perform better which means there is a strong correlation between myopia and good grades, but obviously myopia is not the reason for good grades.
To find the causality between X and Y, we can use counterfactual reasoning (or counterfactual inference). "Counterfactual" means "counter to the facts", and the inference procedure is to image:
Was it X that caused Y? What if I had acted differently? 

In this talk, I will first give you some examples to help you understand the meaning of counterfactual reasoning.
Next, I will talk about three applications using counterfactual reasoning: one natural language understanding task (take sentiment analysis as an example) and two recommendation tasks (one is to tackle clickbait problem and another to mitigate popularity bias).

Hope you will enjoy!

Paper lists: 
[1] Learing the Difference that Makes a Difference with Counterfactually-Augmented Data 
[2] "Click" Is Not Equal to "Like": Counterfactual Recommendation for Mitigating Clickbait Issue
[3] Model-Agnostic Counterfactual Reasoning for Eliminating Popularity Bias in Recommender System

Time: Wed 4:00pm
Venue: SEIEE 3-414
Best wishes.
Sansa