Netflix’s Machine Learning Platform : Computational Causal Inference
Netflix’s Machine Learning Platform : Computational Causal Inference Computational causal inference (CompCI) is an emerging interdisciplinary field that sits at the intersection of causal inference, machine learning, and software design.
Here we introduce CompCI and its goal, and explore the deep synergy between CompCI and machine learning, investigate open challenges, and find out different ways to make CompCI scalable in the business operational settings.
Computational causal inference at its heart is a software engineering framework that exploits the growing relationship between causal inference, online experimentation, and algorithmic decision making. The prime objective of CompCI is to develop a causal inference model that trains well on a massive volume of dataset and gives robust performance on unseen observations.
Symbiotic relationship between CompCI and ML
Machine learning algorithms are being used to drive sales, business operations, enhance recommendation content, and personalize user experiences. These algorithms are tested online in order to determine whether they can positively affect the company, and causal inference serves as an independent and unbiased evaluator for the online experiment. Furthermore, the predictions from the machine learning models can be debiased by utilizing a causal effect technique called inverse propensity weights.
Likewise, CI models have also benefited from ML approaches. For example, when relationships in the data can be modeled as a graph, there is a method, which determines heterogeneity in treatment effects and do-calculus, that establishes a programmatic structure for causal effects.
Algorithmic policy making and experimentation platforms are the two particular areas where we see the strong symbiotic relationship between Machine Learning and Causal Inference.
In policy making, the algorithm that outputs an action to take can be optimally personalized to guarantee the maximum reward function. This decision-making algorithm is then tested in an online experiment that gives the causal effect on key performance indicators due to the new algorithm.
Experimentation Platform (XP)
XP is at the core of how companies enhance the customer experiences. Companies like Uber and Netflix deploy various experimental techniques to understand its user base, different segments in the user base, and how they change over time.
An XP models a variety of causal effects for both online and controlled experiments: common average treatment effect, conditional average treatment effects, and time dynamic treatment effects. The OLS method is used to measure average treatment effects, conditional average effects, and time-evolved effects, with computational complexity Big-O(np^2), where n is the number of observations and p is the number of predictors.
Algorithms Policy Making Engines
Policy algorithms automate the process of decision making by sequentially recommending a set of actions that helps the systems to incrementally reach a better state. For each of the n users, the algorithm decides an action among K distinct actions. Each user has features, x, and each action generates the reward function, R, with respect to key performance indicators.
A deterministic policy function takes x as an input and returns an action that is supposed to produce the optimal reward. Thus the model’s prime objective is to compute the optimal policy function that maximizes its difference with the current policy function. The computational complexity of this framework is the same as that of XP.
The chief constituents of CompCI, like XP and Policy engines, are being used by companies like Uber and Netflix to drive innovation, automation, and personalized experiences.
However, CompCI faces numerous challenges, such as the generalization of CI models, software design, scalability, and numerical computation. The authors of this paper urge fellow research engineers to have a look at the following set of open challenges:
- To structure CompCI around the class of models that are differentiable and train them generically using SGD requires hyperparameter tuning. It is not yet clear how the risk of poor convergence affects the performance of causal effects estimators.
- A software that detects the marginal treatment effects needs to be developed.
- The impact of conditional randomization and the availability of treatments at a given instance of time is unclear.
In conclusion Companies can utilize CompCI paradigms in order to integrate causal effects into large engineering systems.