An Information View Of Causality And Its Applications In Industrial Personalize Incentives With Uplift Modeling

Does exercise help lower cholesterol?There are many difficulties in answering this question in traditional statistics.It is an example of Yule-Simpson’s paradox.The reason for the difficulty is that the question is essentially a causal question,while traditional statistics avoid causation and use pure statistical tools to answer this question.In recent years,the discovery and research of the causal ladder has given us formal and systematic tools to study causality.The study of causality can be traced back to Aristotle more than 2,000 years ago.The Hume question in philosophy is generally speaking:Can human beings obtain the law of causality from limited experience?In the philosophical exploration of causality,in 1999,Collier first formally proposed that causality is information transmission.Illari pointed out in his book Causality:Philosophical Theory Meets Scientific Practice in 2014 that "the information account causality might be useful to the scientific problem of how we think about,and ultimately trace,causal linking,and so to causal inference and reasoning".This article studies causal modeling theory and answers causal questions from the perspective of informational account on causality,especially about causal effect estimation,and explore the application of causal inference in personalized incentive allocation in internet industry.Theoretically,do-calculus operator models causality from the perspective of interventionism.Accordingly,we explore what the operator corresponding to information account of causality,use it to represent causal effect,and study the corresponding operator calculus theory to identify causal effect.Furthermore,we pay attention to the natural relationship between path-specific effects and information transfer along a path,and develop the theory of path intervention and path-specific effect based on informational decomposition of structural causal model;Finally,we return to the essence of modeling theory and point out that a major problem in causal modeling is circular causality.We propose a meta structural causal model from the perspective of information transmission,connect data and causal mechanism,and turn the problem of circular causality theory into a computational problem,which can be solved under certain scenario.In terms of application,we explore the application of causal inference in recommendation system in the industry.Specifically,we study how to allocate personalized incentives for Internet short video platform,in order to improve user activity and watch-time.The first chapter is the introduction.We introduce the research background and development of causality science,as well as the research work and innovation of this thesis.The second chapter is mainly the literature review,which briefly summarizes the research on causality,especially the literature related to the research content of our dissertation.Starting from the third chapter,we introduce our theoretical research results.Firstly,from the perspective of informational causality,we propose info intervention in spirit to represent causality hypothesis and information transparently and verifiably.After the research on info intervention,in Chapter 4,we will study the path-specific effects,propose the path intervention based on the information decomposition of the structural causal model,and point out that path-specific effects can still be formally defined with our method even under the cyclic causal relations,and the identification of our path-specific effect does not required the "no witness recanting" hypothesis for identification.Facing up to the problem of cyclic causality,we propose an active set method to connect data and meta structural causal model,and explore the general modeling theory from the perspective of information account on causality.In Chapter 6,from theory to practice,we study the conditional causal effect estimation method for uplift modeling in industry.We propose a Unified Discriminative Causal Forest(UDCF)modeling multiple treatment sensitivity estimation,give a Dual Gradient Bisection(DGB)algorithm to solve the budget allocation problem,and propose a new evaluation metric for multiple treatment constrained uplift modeling in personalized incentive applications.Finally,in Chapter 7,we summarize the work of the full text,and also mention the doubts and problems found in the research work.Looking to the future,causality is the inevitability in chance,certainty in probability,invariant in change,and the immutability in the simple essence behind the complex phenomena.A fundamental purpose of scientific research is to discover Causality,helping to express human experience and wisdom concisely and gracefully.Powerful causal information processing and utilization capabilities are one the key element for such prosperity of human civilization.We hope that in the future society,causal reasoning will be integrated into every discipline,especially the close integration and improvement of AI,so that it has the same powerful causal information processing capabilities as human beings.