Amin Rahimian, Research

Large-scale sociotechnical systems in the age of data. My research is primarily concerned with the challenges of analysis and decision making in large scale sociotechnical systems such as online social networks. Such systems operate in human (social and economic) as well as machine (dynamical and algorithmic) dimensions that are complex and conjoined. My work at the intersection of networks, data, and decision sciences is motivated by applications that involve sociotechnical networks and the demand for models and techniques that can work with the massive detailed data collected about them.

Scalable and resilient operation of large-scale networks is dependent on our ability to learn their properties from dispersed observations (i.e., distributed inference) and our ability to influence them from localized action points (i.e., decentralized interventions). The main goal of my research is to facilitate these functionalities; first and foremost, by identifying paradigms that best describe and predict individual and aggregate behaviors. Such behaviors are shaped by the interactions among network components (so-called agents) and they depend critically on the type and quality of the information available to each agent. Understanding these interactions helps us determine whether information transmission is efficient, and guides us through interventions that rely on information flow (e.g., adoption of a new technology). My ultimate aim is to demonstrate the power of behavioral paradigms in control and design, where their application leads to improved policies and practices.

I address these challenges (distributed inference and decentralized interventions) through the lenses of network and data sciences, by relying on tools in applied probability, graph theory, applied statistics and machine learning, algorithms and complexity, as well as game and decision theory. These tools allow me to highlight key structural features that influence behavior; features such as presence of influential agents, observational and information asymmetries, data flow topology and heterogeneity.

Most of my early focus has been on technological networks (such as robotic sensor and actuator networks), where I have contributed to problems of decentralized control and distributed estimation: achieving global objectives by injecting local inputs on the one hand, and learning system properties given restricted partial observations, on the other. My current focus is mostly on the sociotechnical system (such as web-based social networks and e-commerce platforms), and I have been interested in both the effectiveness of information transmission, sharing, and exchange through revealed actions, as well as the effectiveness of decision making using the available data. In fact, these issues are inter-related, as are control and estimation. On the one hand, the quality of decision-making depends on the available information. On the other hand, decisions reveal some of the information at disposal of the decision-maker.

Information, Inference, and Intervention Design for Sustainable and Resilient Sociotechnical Systems

Most of my ongoing research is focused on dynamic structural models that facilitate distributed inference and decentralized interventions on web-based sociotechnical systems (e.g., online markets, crowdsourcing platforms, and sharing economy). I am mostly interested in the development of structural models that take into account the behavioral mechanisms that influence the decision-makers in different situations. The latter often requires a deep understanding of social and human sciences to accurately characterize a decision scenario subject to bounded-rationality and cognitive biases. Such models have the advantage of being amenable to policy interpretation while revealing the implications of the theory of mind in a particular scenario. On the one hand, I rely on methods in probability, random processes, and algorithms to execute asymptotic analysis of these models. Such results allow us to reason through policy interventions in the large-scale, as the dimensions of interest (e.g., number of interactions) increase. However, the full utility of these models is not just in the insights that we obtain from their asymptotic analysis. Moving forward, I would like to demonstrate the potential of these models in revealing the statistical and causal dependencies between various variables of interest (observable or latent). Structural models are notorious for not being amenable to inference  —  having complex and often intractable likelihood functions. To overcome this intractability, I rely on approximate Bayesian computation and simulation-based techniques to automate inference, estimation, and prediction on dynamic structural models.

The three processes that I am studying are:

• social contagion
  The spread of new ideas, products, and behaviors over social networks depends on the contact structure among individuals. In the presence of social or biological contagion, decision-makers in marketing, public health, development, and other fields strategize about where and how to intervene in a network (e.g., by giving a free product or suggesting a new contact).

• network formation
  Pervasiveness of social and economic network data give rise to important questions about the formation and evolution of links between the networked entities (agents). A better understanding of the underlying mechanisms of network formation has important implications for all kinds of social policy and business strategy, where network interactions and externalities play a role.

• online reputation
  Reviews and ratings posted on electronic commerce platforms strongly influence the purchasing and browsing behavior of customers, and play an important role in revenue generation.

Two of my most recent works in these areas are the following:

• In [1.1], we consider the choice of k seeds in a social network to maximize the expected spread size. Most of the previous work on this problem (known as influence maximization) focuses on efficient algorithms to approximate the optimal seed sets with provable guarantees, assuming the knowledge of the entire network graph. However, in practice, obtaining full knowledge of the network structure is very costly. To address this gap, we propose algorithms that make a bounded number of queries to the graph structure and provide almost tight approximation guarantees.

• In [1.2], we study how interventions that change the network structure can increase the speed of spread. For simple models in which contagion spreads through each edge independently at random, interventions that randomly rewire the edges would increase the speed of spread. However, for other contagion models that require multiple exposures before adoption (i.e. threshold-based contagions), recent work has argued for the opposite conclusion: highly clustered, rather than random, networks facilitate spread. In [1.2], we characterize the conditions under which we can reverse the latter result by allowing a small probability of sub-threshold adoptions.

See here for a list of the relevant publications.

Privacy of Interconnected Learning Systems and Operations

The data most useful for learning, coordination, and intervention are often the most sensitive, distributed across parties, strategically withheld, or only partially observable. This work asks how organizations and platforms can still estimate, decide, and intervene well under those constraints. It treats privacy as a design principle rather than only a constraint on data use, and in some settings as a design instrument that improves how a system operates.

In [2.1], we argue for building this kind of protection into networked healthcare by design, across the data lifecycle from collection to inference to decisions, so that lawful, trustworthy data sharing becomes the default rather than a barrier to overcome. The annotated reading list Yuxin Liu and I wrote [2.2] offers an entry point to where privacy meets incentives, data acquisition, and operational decisions; it followed a tutorial we organized with Marios Papachristou and Juba Ziani.

• 

  

  learning and inference from private, distributed data
  In [2.3], we study how agents in a network estimate shared statistical quantities from privately held data while protecting both their observations and, in some settings, their local connections. The aggregation rules add calibrated noise and converge in finite time; among mechanisms of this kind, Laplace perturbation gives the best convergence for a fixed privacy budget. The analysis makes precise a trilemma among privacy, communication, and decision quality, and the methods are validated on U.S. power-grid and German household electricity data.

  In [2.4], we extend from continuous estimation to discrete inference and hypothesis testing, with distributed maximum-likelihood protocols applied to multicenter clinical-trial survival analysis and federated genomic biomarker discovery under false-discovery-rate control. Because institutions exchange only privatized summaries rather than patient records, this lets smaller and more diverse hospitals join multicenter studies that data-sharing agreements would otherwise delay or prevent.

• 

  

  privacy, incentives, and strategic information sharing
  In [2.5], we study how people learn from others’ choices when privacy limits what those choices reveal about their private signals. Even without privacy constraints, later participants may follow the crowd rather than their own information, allowing mistaken beliefs to persist through information cascades. Carefully designed privacy-preserving randomization can weaken fragile cascades and improve collective learning. For example, different flip rates under true and false states can help later learners distinguish the two states, accelerating learning despite the added noise.

  In [2.6], we study why competing firms may withhold data even when pooling it improves their forecasts. Privacy noise limits what rivals can learn from a firm’s contribution, softening the competitive pressure created by sharing—but also making the pooled information less useful. External signals supplied by the platform preserve the informational benefit: privacy protection and external signals act as complements to make voluntary data sharing worthwhile when either alone is insufficient.

• 

  

  targeted interventions on sensitive networks
  In [2.7, 2.8, 2.9], we design targeted network interventions when the network is itself sensitive, only partially observed, or too costly to reconstruct, motivated by HIV prevention and PrEP delivery. Partial network information and contact-tracing-style cascade data can still support effective seeding while protecting the people involved.

  The work frames and resolves a second trilemma: the cost of acquiring network information, the benefit of sharper targeting, and the privacy risk of using personal data. Results come from empirically grounded simulations of sexual-activity networks, and they extend our earlier seeding work under costly network information [1.1] to the privacy setting.

• 

  

  private measures of network structure
  In [2.10], we release measures of assortative mixing and connectedness, capturing whether edges tend to link similar nodes and how connected the network is, while protecting both the edges and sensitive node attributes. These indices resist direct privatization because a node's attribute can influence the index across the whole network, so protecting it naively forces large, compounding noise that degrades the estimates.

  We instead use a two-stage privatize-then-debias procedure, adding noise for the guarantee and then correcting the bias it introduces, with consistency and asymptotic normality for the released statistics.

• 

  

  mechanisms shaped by the geometry of the data
  In [2.11], we introduce Conformal-DP, a density-aware mechanism for data lying on curved spaces (Riemannian manifolds), rather than in flat coordinates. Using a conformal transformation, we match the perturbation to the local data density to improve the privacy–utility trade-off for heterogeneously distributed, non-Euclidean data. This opens new opportunities to design privacy noise mechanisms for structured data objects such as networks, manifolds, embeddings, and genomic sequences, whose geometry can be distorted by coordinate-wise noise.

My third Ph.D. student, Yuxin Liu, built his dissertation, Privacy-Preserving Information Design for Learning, Strategic Interaction, and Networked Platforms, on the sequential-learning, information-sharing, private-seeding, and connectedness-index results described above.

Sensitive data must often be shared and analyzed to advance health and social science and to support better policy, while the same data can create risks to privacy, dignity, self-determination, and protection against bias and discrimination. The direction I am most focused on is designing protection around the way people are connected and systems operate, through contagion, social ties, social learning, or shared genetics, rather than treating privacy as a cost imposed from outside. Designed this way, protection can shield sensitive dependencies while preserving the quantities required for inference and operational decisions; for example, improving association tests on network and individual-level data by attenuating spurious dependence, or keeping assignment and matching decisions near-optimal while sensitive attributes stay protected.

See here for a list of the relevant publications.

Public Health Dynamics, Operations and Policy

My research in public policy and population health builds on the work on social contagion and network diffusion described above. My entry into public health came through the COVID-19 pandemic, where I became interested in how policy effectiveness depends not just on where a policy applies but on how its effects travel through social networks, via spillovers between connected populations, behavioral responses, and the diffusion of beliefs. The work since then has developed along two complementary lines: measuring how health risks and protective policies propagate through socio-spatial networks, and building computationally tractable methods for evaluating interventions when the underlying simulations are expensive.

• policies, norms, beliefs, and behavior during COVID-19
  This line of work began during my postdoc at MIT. With Dean Eckles, Sinan Aral, and colleagues, we studied spillovers in mobility and social distancing across socially and geographically connected U.S. regions, showing the cost of uncoordinated policy responses [3.1]. We then fielded a global survey in 67 countries with more than two million responses on COVID-19 beliefs, behaviors, and norms [3.2], and ran a preregistered randomized experiment with nearly half a million participants across 23 countries showing that accurate descriptive-norm information increased intentions to accept a COVID-19 vaccine [3.3].

• social networks and opioid overdose mortality
  In [3.4], we ask whether opioid overdose mortality is shaped by social ties. Social and geographic proximity are strongly correlated, so measuring social influence means separating the two: using Facebook's Social Connectedness Index we build a network-weighted exposure measure and estimate its association with county overdose mortality, controlling for mortality in spatially proximate counties and for demographic and clinical covariates. A one-standard-deviation increase in deaths in social proximity is associated with thirteen additional deaths per 100,000 population in the contiguous United States. The social association remains statistically significant across a range of alternative model specifications (network and spatial autocorrelation models, two-way fixed effects, two-stage least squares), while the spatial association does not hold consistently.

• socio-spatial dynamics of suicide and firearm policy
  In [3.5], we apply the same design to county suicide mortality over 2010–2022. A one-standard-deviation increase in the suicide rate of socially connected counties is associated with about 2.8 additional suicide deaths per 100,000 people. Extreme Risk Protection Orders are state laws that temporarily restrict firearm access for people at risk of self-harm; a one-standard-deviation increase in social exposure to ERPO-implementing states is associated with about 0.2 fewer suicide deaths per 100,000, even where the policy is not locally enacted, and persists under geographic and state-by-year controls. Harmful exposures and protective policy appear to diffuse through the same social ties, which argues for prevention strategies that account for network structure alongside conventional geographic targeting.

• policy evaluation with costly simulations
  Population-scale agent-based models of opioid use disorder are costly to run, which makes systematic policy comparison difficult. In [3.6] we allocate simulation runs across treatment conditions by confidence-interval width, matching uniform sampling's accuracy with far fewer runs, and in [3.7] we identify which estimator to prefer at a given budget. In [3.8] we show Gaussian process regression can recover how two diseases spread differently across space from simulated populations alone. In [3.9] we pair a Gaussian process that learns how treatment effects vary across counties with a response function that maps intervention levels to mortality, and a two-step sequential design that chooses which county and intervention combination to simulate next; for Pennsylvania counties this reaches under 5% average relative error using fewer than 2% of the runs an exhaustive design would require.

• county-level heterogeneity in harm reduction and treatment
  In [3.10], we calibrate a Markov model of opioid use disorder separately to six Pennsylvania counties spanning urban, intermediate, and rural settings, and project overdose mortality for 2025–2029 under separate and combined increases in naloxone and buprenorphine dispensing relative to each county's projected baseline. A 30% increase in naloxone dispensing is projected to reduce the 2029 overdose death rate by roughly 70% in Allegheny County but about 14% in Erie, compared with the no-increase baseline. The intervention with the larger projected benefit also varies by county: naloxone produces a larger reduction in Allegheny, while buprenorphine produces a larger reduction in Erie, supporting locally tailored harm reduction and treatment planning.

My first Ph.D. student, Kushagra Tiwari, developed the socio-spatial mortality studies in his dissertation, Socio–Spatial Network Dynamics of Public Health Crises in the US. My second Ph.D. student, Abdulrahman Ahmed, developed the simulation-based evaluation methods in his dissertation, Computationally Efficient Simulation-Based Policy Evaluation for Epidemic Interventions: Metamodeling, Sequential Design, and Localized Effect Estimation.

See here for a list of the relevant publications.

Group Decision Making, Social Learning, and Collective Intelligence

In my Ph.D. dissertation, I studied the purely informational interactions of individuals in a group, where they receive private information and act based on that information while repeatedly observing each other's beliefs or actions. Such situations arise, for example, in jury deliberations, expert committees, and medical diagnoses. I developed parallel theories of group decision making following Bayesian and non-Bayesian approaches. In the Bayesian setup, I characterized computational barriers and investigated special structures in which computations simplify. In the non-Bayesian framework, I proposed the no-recall model to reconcile existing heuristics for belief formation and decision making in groups. An overview of these foundations appears in [4.1].

The following are some highlights from my research on group decision making and social learning:

• computational limits of Bayesian learning
  Characterizing the hardness of Bayesian group decisions and reasoning in opinion-exchange networks, where agents infer from private information and histories of others' actions or beliefs [4.2, 4.3].

• learning without recall
  Explaining how simple belief-update rules arise from locally Bayesian inferences that attribute others' actions to private signals without accounting for the full history of interactions [4.4–4.6].

• foundations of linear opinion dynamics
  Deriving familiar rules for updating opinions from a model of heuristic inference and decision making [4.7].

• switching to learn
  Combining social and individual updating to learn while limiting communication [4.8].

• cognitive biases and repeated exposure
  Studying how hindsight bias impedes learning and how repetition shapes beliefs when recall is limited [4.9, 4.10].

Two of the main issues that arise in the study of decision-making organizations are information aggregation and decision-flow architecture. My formal theories of group decision making provide structural insights into both issues and can be operationalized to avoid redundancy and increase the efficiency of group decisions among heuristic decision-makers. For example, carefully scheduled exchanges can aggregate dispersed information without repeatedly counting the same evidence.

The next two directions examine these issues when agents learn sequentially from others' actions or repeatedly exchange estimates or beliefs with their neighbors.

• 

  

  sequential learning and information cascades
  With Yuxin Liu, I study how private signals and others' observed actions shape belief accumulation and information cascades, and how structured randomization can accelerate learning with continuous signals, even when introduced to protect privacy [4.11].

• 

  

  distributed estimation and learning
  In [4.12], we study how local exchanges over heterogeneous networks aggregate observations that differ in quantity, quality, and timing, with convergence to globally efficient estimators and finite-time performance guarantees. With Marios Papachristou, I extend this framework to differentially private distributed estimation and collective inference, with explicit bounds on accuracy and convergence under privacy constraints [4.13, 4.14].

The privacy mechanisms and applications of these two directions are discussed above.

Whether social information improves collective performance depends on the task and on how communication is structured. We examine these factors in the following:

• 

  

  

  task-dependent collective intelligence
  We show how the distribution of initial estimates moderates the effect of social influence on collective accuracy [4.15]. Through experiments and simulations, we also study how algorithmic mediation and rewiring of communication networks change group performance [4.16, 4.17].

• 

  

  social learning in collective search
  In [4.18], we study how information sharing shapes group foraging. Intermediate social learning ranges balance independent exploration with coordinated exploitation, while broader sharing can improve equity at the expense of efficiency and temporal stability.

Algorithmic, computational, and optimality aspects of decision-making in organizations are vastly unexplored. There are many untapped potentials for applying these techniques to improve the operations of teams in medical, legal and other industrial decision-making organizations. By investigating the effects of heuristics and biases, we can improve the practice of social and organizational policies, such that new designs can accommodate commonly observed biases, and work well in spite of them. My ongoing work extends this agenda to human–AI collaboration and collective decision making among AI agents, examining how task-dependent complementarities and communication shape decision quality and the development of human expertise.

See here for a list of the relevant publications.

Control, Analysis, and Design of Networks

Failure detection and isolation in directed networks [5.1].

Networked dynamic systems build upon complex interactions between many co-evolving sub-components. Their industrial applications are plentiful and diverse ranging from chemical and biological processes to robotics and the power grid. My results touch upon many aspects of the analysis and design for such complex systems, including distributed detection, estimation, reliable design, and decentralized control. I have been particularly interested in the following questions:

1. How can we identify failures and structural variations from the observed dynamics of a limited number of output nodes?

2. How can we exercise control over the entire networks by accessing only a few of their input nodes?

3. How can we learn about environmental variables from the dispersed and disparate observations of different nodes in a network?

Tools from control theory, optimization, signal processing, statistics, graph theory, combinatorics, and discrete algorithms have enabled us to propose effective solutions to each of these problems.

Some highlights from our results on the analysis and design of networked dynamic systems are as follows:

• Efficient Placement of sensors for detection and isolation of failures (submodular maximization) [5.1]

• Detection and isolation of link failures based on the observed outputs (linear systems theory) [5.1]

• Distinguishing network structures based on their nodal dynamics (algebraic graph theory) [5.2]

• Protecting structural controllability against link and node failures (graph theory, combinatorics) [5.3]

• Efficient methods to rank links and nodes for the preservation of structural controllability (graph theory, combinatorics) [5.3]

• Efficient placement of actuators for optimal control performance (submodular maximization) [5.4,5.5]

• Learning from intermittent streams of heterogeneously correlated data (distributed estimation, learning theory, Bayesian statistics) [5.6]

Semicircle spectrum of a random network [5.9].

Some of my more recent works in this area are focused on developing tools that pioneer applications of random matrix theory in the modeling and analysis of large-scale networks and dynamic systems. Such tools have great potential for addressing questions whose depth and analytical complexity have impeded researchers so far. They are also very versatile and expressive when it comes to modeling and explaining structural features of very large-scale systems such as robotic swarms or the extremely high dimensional data sets that arise in the study of the Internet and online social networks. Some of my main results so far are:

• Limiting spectral moments of random matrices with rank one pattern of variance [5.7]

• Limiting spectral moments of random graphs with specified expected degrees (the Chung-Lu model) [5.8]

• Theoretical justification for the quasi-triangular spectrum of power-law networks (reported in empirical literature) [5.8]

• H2 norm and the spectrum of the controllability Gramian of random linear systems [5.9]

See here for a list of the relevant publications.

Controller Design and Tuning

The versatility and simplicity of proportional, integral and derivative (PID) actions have made them the most common form of feedback control, and almost universal all across the agriculture, transport, chemical and manufacturing industries, or even telecommunications and web services. Our ability to efficiently tune these controllers is key to their widespread adoption. My collaborators and I have advanced the state of the art in process control by proposing new ways of tuning feedback controllers. Our methods focus on explicit characterization of feasible domains in the control parameter space. Tools from complex analysis allow us to study the root boundaries of various polynomials that arise in the study of the feasible domains.

Centroid tuning with a phase margin (PM) constraint [6.3].

Fractional order control builds on a natural generalization of the integral and derivative feedback actions based on fractional order differential operators. We have successfully applied several of our techniques in the design and stabilization of fractional order controllers. Here are some of the highlights from our results on the design and tuning of feedback controller:

• Constrained tuning and parameter optimization over the feasible regions [6.1, 6.2]

• Using the stability region centroids for robust and non-fragile stabilization [6.3,6.4]

• Stability regions for fractional-order controllers and their integer approximations [6.5]

Although I no longer follow this line of work, the study of feedback control early in my research career has influenced my understanding of large-scale complex systems.

See here for a list of the relevant publications.