Source-linked AI summary
Causal Embeddings for Recommendation
Stephen Bonner, Flavian Vasile
TL;DR
Recommendation systems often predict organic behavior even when their purpose is to change it, creating a gap between evaluation and intervention objectives. The paper introduces CausE, which learns randomized-exposure outcomes by adapting from large biased-feedback data and small randomized-feedback data. CausE significantly outperforms classical matrix factorization and recent causal recommendation methods.
Problem
Standard recommendation evaluation predicts observed behavior rather than the individualized causal effect of recommendations relative to a control policy.
Method
CausE modifies matrix factorization to learn from a large biased-exposure sample and a small unbiased-exposure sample, regularizing discrepancies between their item representations.
Results
CausE significantly outperforms classical matrix factorization and recent causal approaches across both datasets.
Takeaways & Limitations
The method is promising for industrial recommendation because real-world feedback is usually biased and matrix-factorization systems require little modification.
Takeaways & Limitations
The current approach does not yet model user sequences; extending it to organic and influenced activity sequences is future work.
Abstract
from arXiv · showhide
Many current applications use recommendations in order to modify the natural user behavior, such as to increase the number of sales or the time spent on a website. This results in a gap between the final recommendation objective and the classical setup where recommendation candidates are evaluated by their coherence with past user behavior, by predicting either the missing entries in the user-item matrix, or the most likely next event. To bridge this gap, we optimize a recommendation policy for the task of increasing the desired outcome versus the organic user behavior. We show this is equivalent to learning to predict recommendation outcomes under a fully random recommendation policy. To this end, we propose a new domain adaptation algorithm that learns from logged data containing outcomes from a biased recommendation policy and predicts recommendation outcomes according to random exposure. We compare our method against state-of-the-art factorization methods, in addition to new approaches of causal recommendation and show significant improvements.
1 INTRODUCTION
Recommendation systems are typically evaluated by predicting organic behavior, although their intended role is to change behavior toward a desired outcome. This paper formulates recommendation as an individualized causal policy problem and introduces CausE to learn randomized-exposure outcomes from biased and randomized feedback.
- 1 INTRODUCTION: Current recommender models frame recommendation as pairwise distance or next-item prediction, using metrics such as MSE, AUC, Precision@K, and NDCG.These formulations evaluate coherence with observed behavior rather than intervention outcomes.
- 1 INTRODUCTION: Recommendations are interventions intended to influence user behavior toward outcomes such as purchases, viewing, or applications.The desired change is measured against a control or logging policy, including a system with no recommendations or a baseline recommender.
- 1 INTRODUCTION: The Individual Treatment Effect measures a policy’s reward for a user-product pair relative to the control policy.The paper seeks a policy maximizing the sum of these individualized effects.
- 1 INTRODUCTION: CausE modifies matrix factorization by using a small sample of randomized recommendation outcomes to create user and product representations.It transfers evidence between a large biased-exposure sample and a small randomized-exposure sample.
- 1 INTRODUCTION: IPS-based estimators struggle when treatment and control exposure probabilities differ substantially, while uniform logging is impractical because it reduces recommendation quality.The proposed strategy instead learns outcomes under randomized exposure from biased data and a small randomized sample.
- 1 INTRODUCTION: CausE extends prior work on biased control plus randomized treatment, generalizes earlier approaches, and is benchmarked against factorization and IPS-based causal methods.The evaluation analyzes ways to use the exploration sample for unseen data from the randomized treatment policy.
2 RELATED WORK
Related work treats causal recommendation through propensity scoring, counterfactual risk minimization, bandit feedback, domain adaptation, and transfer learning. Existing recommendation methods also debias matrix factorization by modeling exposure and reweighting observed feedback.
- 2 RELATED WORK: Inverse Propensity Scoring reweights observational outcomes to estimate a target policy’s reward as though data came from a randomized trial.Clipped IPS addresses variance problems when treatment and control policies differ substantially.
- 2 RELATED WORK: Double Robust estimation combines direct-method and IPS estimates and is consistent when at least one model is well specified.This provides an alternative estimator for evaluating new policies from logged data.
- 2 RELATED WORK: Counterfactual Risk Minimization and Batch Learning from Bandit Feedback address performance under interventions that change the data distribution.POEM optimizes a variance-regularized clipped-IPS estimator, while BanditNet trains neural policies from contextual bandit feedback.
- 2 RELATED WORK: Domain-adaptation approaches learn representations on factual data and predict counterfactual outcomes, while transfer-learning methods combine A/B-test traffic with pre-A/B control data.Both lines of work use distributional or historical control information to improve causal-effect prediction.
- 2 RELATED WORK: Causal recommendation methods such as ExpoMF model exposure and use inverse exposure probabilities to weight implicit-feedback data.The resulting weighted data approximates an experiment in which users are randomly shown items and supports domain adaptation.
3 PROPOSED APPROACH
The proposed approach predicts recommendation outcomes under random exposure by jointly learning from large biased-control data and small randomized-treatment data. It adapts representations for users and products across control and treatment policies, with regularization linking the tasks.
- Data and objective: The method targets prediction of recommendation outcomes under random exposure using a large logging-policy sample and a small randomized-policy sample.The randomized sample may come from uniform exposure or an exploration strategy, while the logging data provide broader coverage.
- Exposure policies with fixed users: For fixed users, the ITE is approximated as the difference between treatment and control inner products for each user-product pair.Product representations differ between policies, while user representations remain fixed in this formulation.
- Joint objective: The treatment task factorizes observed randomized-policy rewards, while the control task leverages the larger logging-policy sample.The objective jointly models treatment and control observations with separate representation parameters and regularization terms.
- Exposure policies with varying users and treatments: The full CausE objective allows both user and product representations to change between treatment and control policies.This formulation addresses exposure changes involving different users and products and recovers the fixed-user case when user-difference regularization becomes infinite.
- Optimization: CausE is optimized with stochastic gradient descent using momentum and a linearly decaying learning rate.The algorithm iterates over batches, looks up user and product indices, and updates control and treatment vectors.
4 EXPERIMENTS
The experiments evaluate recommendation outcome prediction under biased control data plus limited treatment-policy data, using standard, propensity-based, and causal factorization baselines. On distribution-shifted SKEW datasets, CausE significantly outperforms competing methods, while performance converges when training and test distributions are constant.
- Evaluation setup: The evaluation predicts treatment-policy outcomes using large control-policy samples and small treatment-policy samples, measuring MSE, NLL, and AUC.The setup simulates limited exploration or A/B-test data alongside extensive logged recommendations.
- Evaluation setup: Four adaptation setups vary whether training uses control data, treatment data, their blend, or an averaged treatment product.The setups are no adaptation, blended adaptation, test-only adaptation, and average test adaptation.
- Datasets: The study uses MovieLens10M and Netflix for evaluation, plus MovieLens100K to examine how injected test data changes performance.The datasets are binarized and transformed to simulate causal effects under uniform exposure.
- Results: CausE significantly outperforms all baselines across both datasets, with CausE-prod-C outperforming WSP2V-blend by 21% MSE and 20% NLL lifts on MovieLens10M.CausE-avg and CausE-prod-C also outperform BanditNet, while BN-blend is the closest competing method.
- Results: On constant-distribution REG data, all methods perform similarly, with MovieLens10M approaches achieving an MSE lift of +8%(±0.02).The reported REG result is omitted from Table 2 for brevity.
- Results: On SKEW data, standard matrix factorization performs badly, while test-only training remains limited by sparsity and blended supervised adaptation improves only marginally.The relevant standard methods are SP2V and BPR; SP2V-test trains exclusively on treatment-distribution data.
5 CONCLUSIONS
The paper introduces CausE, a matrix-factorization method that optimizes causal recommendation outcomes by adapting biased exposure data with a small uniformly exposed sample. It reports clear improvements over classical and recent causal approaches and argues that the method is practical for existing matrix-factorization systems, while identifying sequence modeling as future work.
- CausE learns causal recommendation outcomes from a large biased-exposure sample and a small uniformly exposed sample.It adds a regularizer matching item representations fitted to the biased and uniform-exposure samples.
- CausE clearly outperforms classical matrix factorization and recent causal methods, including IPS-weighted factorization and BanditNet.
- The approach fits industrial settings because most feedback is collected under biased recommendation systems and many recommenders use matrix factorization.The authors state that existing matrix-factorization systems would require little modification to experiment with CausE.
- Future work will extend the approach to user sequence modeling using both organic and influenced activity sequences.