What is the impact of a recommender system? In a typical three-way interaction between users, items and the platform, a recommender system can have differing impacts on the three stakeholders, and there can be multiple metrics based on utility, diversity, and fairness. One way to measure impact is through randomized A/B tests, but experiments are costly and can only be applied for short-term outcomes. This talk describes a unifying framework based on causality that can be used to answer such questions. Using the example of a recommender system's effect on increasing sales for a platform, I will discuss the four steps that form the basis of a causal analysis: modeling the causal mechanism, identifying the correct estimand, estimation, and finally checking robustness of the obtained estimates. Utilizing independence assumptions common in click log data, this process led to a new method for estimating impact of recommendations, called the split-door causal criterion. In the later half of the talk, I will show how the four steps can be used to address otherw questions such as selection bias, missing data, and fairness questions about a recommender system.
What is the impact of a recommender system? The truth obviously lies somewhere in the middle. Both are exaggerated.
Suppose you are Amazon and you are While the concepts are general, they are best understood through an example. Causal: how much activity Suppose you want to improve recommendation. One of the metrics you want is for novel recommendation
And Ideally, we would want such an estimate for every product. And in many cases, infeasible. E.g. considerable effect on user experience.
Question: rec has value Question: can randomize order. Or show random recommendations: why costly? Answer: can do but we need offlne metric..can be used to train new algorithm.
But if you just think about it, obs. CTR is almost surely an overestimate. It is helpful to think about in terms of causal and convenience. By design, a recommender system shows similar products,
In our case, it is page visits due to recommender and direct visits.
Story: yd is instrument. Not coming automatically but more validating. Say and this actually happened..oprah invited road book.
Everything that is affecting pr outcome should affect auxiliary. Can think of as giving us exclusion. But more broadly, serves to remove this arrow.
Observed effect is also the causal effect.
But we can actually do more general.
Improve quality of image.
All products is it method? Baseline: A method that can generate valid instrument
Can discover those that we would not think of.
Causal inference in Recommender Systems
Causal Inference in
Senior Researcher, Microsoft Research India
Invited Talk: REVEAL Workshop @ACM RecSys 2020
How to evaluate a recommender system?
• Is the predicted rating similar to a user’s
• Does the user click on a
• Does the system exclude certain items
• Does the system recommend items
different from each other?
Insufficient for the questions we
want to answer.
Does the recommender system increase
Does it shape what people buy or
Does it create “echo chambers” or
make people more polarized?
Simple example: The “Harry Potter” Problem
Suppose a recommender always recommends the next book by the same author.
High accuracy and high coverage system. Diversity can also be high if user reads diverse genres of books.
Harry Potter 2
By J.K. Rowling
By Cormac McCarthy
A causal view of a recommender system
Key question: What would be the outcome metric in a world without the
Causal effect of
𝑃 𝑂𝑢𝑡𝑐𝑜𝑚𝑒 𝐝𝐨(𝑅𝑒𝑐))
𝑃 𝑂𝑢𝑡𝑐𝑜𝑚𝑒 𝐝𝐨(𝑅𝑒𝑐 = 1)) 𝑃 𝑂𝑢𝑡𝑐𝑜𝑚𝑒 𝐝𝐨(𝑅𝑒𝑐 = 0))
Causal Impact of Recommender= 𝑃 𝑂𝑢𝑡𝑐𝑜𝑚𝑒 𝐝𝐨(𝑅𝑒𝑐 = 1)) − 𝑃 𝑂𝑢𝑡𝑐𝑜𝑚𝑒 𝐝𝐨(𝑅𝑒𝑐 = 1))
Comparing to a counterfactual world provides
new, causal metrics
Serendipity: “Recommendation helps the user find a surprisingly interesting item
they might not have otherwise discovered” ---Herlocker et al. 2004 (TOIS)
But so far we lacked the tools to measure such metrics.
Increase in Clicks
Fairness by Parity
Fairness by Equal
Today’s talk: How to estimate causal metrics
for a recommender system?
1. Case study: Estimate the impact of Amazon’s recommendation engine
Describe the four steps of causal analysis:
1. Model causal mechanisms in a system.
2. Identify the correct metric to estimate.
3. Estimate the metric.
4. Check robustness of the estimate to unobserved confounding.
2. New, causal metrics: How a causal inference view enables us to ask new
questions about a recommender system?
DoWhy: A Python library for causal inference that implements the four
Causal Impact: How many additional views does a
recommender system bring?
Increase in Clicks
Hypothetical experiment: Randomized A/B test
Can we develop an offline metric? 9
Treatment (A): Observed Control (B): Counterfactual world
Step 1: Modeling the causal mechanism and
identifying the confounding factors
Visits to The
Rec. visits to
for Old Men
No Country for
Observed activity is almost surely an
overestimate of the causal effect
Step 2: Identification--Is there a way to
recover the causal effect from observed data?
To remove convenience clicks, need
a proxy for unobserved demand.
“Backdoor criterion”: 𝐄 wY/X
where the weight
𝑤 = 1/𝑃(𝑋 = 1| 𝑈𝑠𝑒𝑟𝐶𝑜𝑛𝑡𝑒𝑥𝑡) captures
demand of the user. (inverse propensity
But method depends on accurately
capturing unknown user context.
Finding a demand proxy using natural experiments:
Split outcome into recommender (primary) and direct visits
All visits to a
Auxiliary outcome: Proxy
for unobserved demand
for recommended product
to Y (𝒀 𝑹)
to Y (𝒀 𝑫)
Example: Product X’s visits change but the direct visits
to recommended product Y are constant (Accept)
Example: Products visits change and direct visits to
recommended product also change similarly (Reject)
Leads to the “split-door” criterion
Criterion: Observed visits through a recommended link are causal only
when 𝑿 ∐ 𝒀 𝑫 .
Visits to focal
More formally, the criterion is based on do-
calculus over the causal graph
Step 3: Estimation with Amazon.com logs
from the Bing toolbar
Out of which 20 K products have at least 10 visits on any one day
Implementing the split-door criterion
< 𝑋, 𝑌𝐷 >
𝑡 = 15 days
Estimate the metric over valid split-door pairs
Using the split-door criterion, obtained 23,000
natural experiments for over 12,000 products.
(~half of all products~20k)
Step 4: Check robustness of the estimate to
What if there is an
that affects the
throughs but not the
• Select plausible values
for the confounder
• Simulate how robust
the estimate is.
Summary: Same process of causal analysis can be
applied to develop metrics for new problems
• Does a system provide same accuracy/performance across
• Rishabh Mehrotra, Ashton Anderson, Fernando Diaz, Amit Sharma, Hanna Wallach, Emine
Yilmaz (WWW 2017). Auditing Search Engines for Differential Satisfaction Across
• How to measure long-term outcomes due to a system that cannot be
measured by randomized experiments?
• If you have a new product, which people to send the
recommendation to such that number of purchases is maximized
(limited budget to send recommendations)?
• Email for a copy.
• Try DoWhy, a Python library for causal inference that implements the four steps
of causal analysis
• Upcoming book on Causal Inference in ML systems (w/ Emre Kiciman):
• Sharma, Amit, Jake M. Hofman, and Duncan J. Watts. "Estimating the causal impact of
recommendation systems from observational data." Proc. ACM EC 2015.
• Sharma, Amit, Jake M. Hofman, and Duncan J. Watts. "Split-door criterion: Identification
of causal effects through auxiliary outcomes." The Annals of Applied Statistics (2018).
Amit Sharma, Microsoft Research India