Follow the Herd: A Data-Driven Approach for Actionable Recourse

Nishtha Madaan; Srikanta Bedathur

cs.AIOct 27, 2026

Follow the Herd: A Data-Driven Approach for Actionable Recourse

Authors: Nishtha Madaan, Srikanta Bedathur

Organizations: Indian Institute of Technology Delhi, India

Abstract

As machine learning systems make increasingly greater decisions about people's lives, it is critical to provide users with actionable recourse about how they can achieve their desired outcomes. This problem is currently tackled by relying on expensive human supervision which involves hand-designing a cost function before searching for low-cost actions that achieve the desired outcome. In this paper, we propose ``Data-Driven Action Recourse" (DDAR), the first approach for providing action recourse without requiring costly supervision from humans about action costs. We build upon the notion that highly likely trajectories in the historical user data should be associated with lower costs. With this, DDAR first builds a dynamics model of historical user trajectories. Under this model, it then searches for trajectories that achieve the desired outcome. In experiments, we find that this simple approach recommends plausible action recourse without requiring human supervision, achieving lower action costs than the existing approaches.

Figures & tables

Figure 1: Overview. In this figure, we illustrate the space of user states. The intensity of blue illustrates how likely the user state is in the population. The grey arrows denote the most likely course or direction that users pursue at the given user state. We equate the most likely direction with the direction of a low-cost recourse. The shaded green area denotes the region of state space that achieves the desired outcome from the decision-making system. Left: We illustrate that the original formulation of Wachter et al. ( Wachter et al. 2017 ) aims only for the nearest recourse, disregarding the actionability and cost of the recourse. Center: We illustrate that although the formulation of Madaan et al. ( Madaan and Bedathur 2024 ) emphasizes plausibility when recommending recourse, however, it ignores the cost to the user performing the recourse. Right: In contrast to both these previous paradigms, our framework takes into account the recourse cost conditioned on the current state of the individual when recommending a recourse. Furthermore, the notion of the recourse cost is acquired in a data-driven manner without relying on costly human supervision.

Figure 2: Architecture. In DDAR, the input user state is first encoded and then passed to a denoising diffusion model to produce a recourse embedding that is decoded to produce the final recourse. Crucially, the denoising iterations are intercepted by a gradient-update step with respect to the decision model, to guide the denoising process towards achieving the desired label.

$t$	Age	Edu	Occupation	Cash	Bank Bal	Life	Cred	Wt	Race	Sex	Label
0	18	0	stud_bachelor 1	-58K	-57K	L	300	46.05	C	M	Loan Denied
1	19	0	stud_bachelor 2	-59K	-116K	L	300	44.12	C	M	Loan Denied
2	20	0	stud_bachelor 3	-57K	-174K	L	300	43.57	C	M	Loan Denied
3	21	1	engineer 0	37K	-136K	T	300	41.68	C	M	Loan Denied
4	22	1	engineer 1	38K	-97K	T	300	40.52	C	M	Loan Denied
5	23	1	engineer 2	42K	-55K	T	300	42.31	C	M	Loan Denied

Table 1: LoanSynth Dataset Example. We provide a sample trajectory of a user in our proposed synthetic LoanSynth dataset. Each row corresponds to a year in the user’s life. Each column corresponds to an attribute of the user state. The last column corresponds to the decision-maker’s label about whether this user would be approved for a loan or not.

$t$	Attempt	College	GPA	GRE Score	# Exp	# LORs	SOP	Gender	Label
0	$1^{\text{st}}$	Tier-3	3.0	Good	2	0	Good	Female	Not Admitted
1	$2^{\text{nd}}$	Tier-3	3.0	Good	3	1	Poor	Female	Not Admitted
2	$3^{\text{rd}}$	Tier-3	3.0	Good	3	1	Good	Female	Not Admitted
3	$4^{\text{th}}$	Tier-3	3.0	Good	3	1	Good	Female	Not Admitted
4	$5^{\text{th}}$	Tier-3	3.0	Good	3	1	Excellent	Female	Admitted

Table 2: GradSynth Dataset Example. We provide a sample trajectory (with

t

as the time index) of a user in the GradSynth dataset. Each column corresponds to an attribute of the user state. The teal-colored columns are those that can be directly intervened upon by the user. The last column corresponds to the label about whether this user would be admitted to college or not.

$t$	Age	Education	Marital Stat.	Occupation	Relationship	Race	Sex	Hrs/Wk	Nat	Label
0	24	Some-college	Never-married	Tech-support	Not-in-family	W	F	37	US	0
1	25	Some-college	Never-married	Exec-manager	Not-in-family	W	F	37	US	0
2	26	Some-college	Never-married	Exec-manager	Not-in-family	W	F	37	US	0
3	27	Assoc-voc	Never-married	Exec-manager	Not-in-family	W	F	37	US	1

Table 3: AdultIncome++ Dataset Example. We provide a sample trajectory (with

t

as the time index) of a user in the AdultIncome++ dataset. Each column corresponds to an attribute of the user state. The teal-colored columns are those that can be directly intervened upon by the user. The last column corresponds to the label about whether this user has an income greater than $50K or not.

Figure 3: We plot the recourse rate as a function of maximum allowable cost for our method and compare it against the baselines on the datasets LoanSynth, GradSynth and AdultIncome++.

Table 7

Figure 4: Recourse rate vs. steps

K

. Left: LoanSynth, Centre: GradSynth, Right: AdultIncome++.

Method	Age	Edu	Occ	Cash	Bank Bal	Spend	Debt	Cred	Race	Sex	Loan
Input	29	1	(’sw_dev’, 8)	30K	83K	L	629.36	75.06	C	F	No
Ours	30	1	(’sw_dev’, 9)	22K	127K	L	775.59	73.90	C	F	Yes
DICE	29	4	(’scientist’, 3)	30K	71K	L	627.35	74.83	C	F	Yes
SCD	36	2	(’sw_dev’, 0)	3K	325K	L	849.99	79.58	H	F	Yes
FACE	42	1	(’sw_dev’, 20)	23K	602K	T	849.99	90.88	H	M	Yes
Wachter	29	4	(’scientist’, 3)	145K	71K	L	627.35	74.83	C	F	Yes

Table 6: Qualitative Comparison of Methods on LoanSynth Dataset. We highlight all changes made by a method in cyan color. We also show immutable attributes which should not change like race and gender, and highlight the ones that were wrongly changed in red color.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Age	Edu	Occ	Cash	Bank Bal	Life	Cred	Wt	Race	Sex	Label
28	1	soft_dev 1	28K	60K	T	470.80	58.12	B	F	No
38	2	manager 0	24K	353K	M	849.99	61.55	B	F	Yes
38	1	soft_dev 11	60K	550K	T	849.99	54.90	B	F	Yes
24	1	stud_mast 1	-58K	-266K	L	300.00	82.90	C	F	No
34	3	artist 4	15K	-350K	M	300.04	82.43	C	F	Yes
34	2	engineer 9	60K	99K	M	616.17	87.18	C	F	Yes

Appendix

Table 7: Qualitative Analysis of DDAR with

K

-step recourse on LoanSynth dataset. We provide input-output pairs from our system DDAR. Each row corresponds to a user’s data at a point in time. Each column corresponds to an attribute of the user’s state.

K=10

Hyperparameter	Value
Learning Rate	0.0001
Learning Rate Warmup Steps	30000
Learning Rate Decay Half-Life	250000
Gradient Clipping	0.05
Diffusion Steps	2000
Batch Size	120

Appendix

Table 8: Hyperparameters of the Diffusion Model.

Method	Age	Edu.	Occ.	Sav. (k)	Inc. (k)	Spd.	Debt (k)	CrSc.	Race	Sex	Ln.
Input	29	1	SD-8	30.4	83.3	lav.	0.63	75.1	C	F	No
DSCD	30	1	SD-9	21.8	127.5	lav.	0.78	73.9	C	F	Yes
DICE	29	4	SC-3	30.3	71.2	lav.	0.63	74.8	C	F	Yes
SCD	36	2	SD-0	2.8	325.1	lav.	0.85	79.6	H	F	Yes
FACE	42	1	SD-20	22.8	602.1	thr.	0.85	90.9	H	M	Yes
Wac.	29	4	SC-3	145.2	71.2	lav.	0.63	74.8	C	F	Yes

Appendix

Table 9: Qualitative Comparison of Methods. Changes in cyan, invalid immutable changes in red.

Method	College	GPA	GRE	Exp.	Rec.	SOP	Gender	Admission
Input	Tier-3	4.0	Good	1.0	0.0	Average	Male	False
DSCD (ours)	Tier-3	4.0	Good	1.0	0.0	Excellent	Male	True
SCD	Tier-3	4.0	Poor	8.0	0.0	Good	Female	True
DICE	Tier-1	4.0	Excellent	8.0	6.0	Excellent	Male	True
Wachter	Tier-1	4.0	Poor	8.0	6.0	Excellent	Male	True
FACE	Tier-3	4.0	Poor	7.0	4.0	Average	Male	True

Appendix

Table 10: Comparison of Different Methods on GradSynth Dataset. We show highlight all changes made my different algorithms in cyan color. We also show immutable attributes which should not change in a plausible recourse like gender, tier of institute, college gpa and highlight the ones that were changed in red color.

Age	Edu	Occ	Cash	Bank Bal	Life	Cred	Wt	Race	Sex	Loan?
29	1	soft_dev 8	30,389	83,330.26	lavish	629.36	75.06	C	F	No
30	1	soft_dev 9	21,816.25	127,471.26	lavish	775.59	73.90	C	F	Yes
33	0	stud_bach 3	-29,344	-88,282.86	thrifty	523.62	54.60	B	M	No
34	1	soft_dev 0	28,183.54	-68,442.25	thrifty	849.99	56.80	B	M	Yes
29	1	soft_dev 2	33,652	93,366.83	thrifty	597.82	58.49	B	F	No
30	1	soft_dev 3	40,846.97	127,471.26	thrifty	775.59	56.80	B	F	Yes

Appendix

Table 11: Qualitative Analysis of System DDAR. We provide input-output pairs from our system DDAR. Each row corresponds to a user’s data at a point in time. Each column corresponds to an attribute of the user’s state.

Method	College	GPA	GRE	Exp.	Rec.	SOP	Gender	Admission
Input	Tier-2	2.0	Average	2.0	0.0	Good	Female	False
Ours	Tier-2	2.0	Average	2.0	6.0	Excellent	Female	True
SCD	Tier-3	4.0	Poor	8.0	6.0	Average	Female	True
DICE	Tier-1	4.0	Poor	8.0	6.0	Excellent	Female	True
Wachter	Tier-1	4.0	Poor	8.0	6.0	Excellent	Female	True
FACE	Tier-3	4.0	Poor	7.0	4.0	Average	Female	True

Appendix

Table 12: Qualitative Comparison of Methods on GradSynth Dataset. We highlight all changes made by a method in cyan color. We also show immutable attributes which should not change in a plausible recourse like gender, tier of institute, college GPA and highlight the ones that were wrongly changed in red color. We see that other baselines often make such mistakes while ours does not.

Algorithmic Recourse

Algorithmic recourse studies actionable changes that can help people obtain favorable outcomes from automated decisions.

See all 26 papers

Optimal Recourse Summaries via Bi-Objective Decision Tree Learning
May 8, 2026Ioannis Chatzis, Jason Liartis, Athanasios Voulodimos +1Algorithmic Auditing Decision Tree Learning
Artificial Intelligence and Learning Systems Laboratory
DRACULA: Hunting for the Actions Users Want Deep Research Agents to Execute
Apr 26, 2026Nishant Balepur, Malachi Hamada, Varsha Kishore +9Deep Research Agents LLM Agent Evaluation
University of Maryland · New York University · Allen Institute for Artificial Intelligence (Ai2)
Algorithmic Recourse Under Competition
Sep 30, 2026Shahin JabbariAlgorithmic Recourse
Drexel University

May 8, 2026J/K move · Enter open · S save

Optimal Recourse Summaries via Bi-Objective Decision Tree Learning

Ioannis Chatzis, Jason Liartis, Athanasios Voulodimos, Giorgos Stamou

Artificial Intelligence and Learning Systems Laboratory

Algorithmic Auditing Decision Tree Learning Counterfactual Reasoning Multi-Objective Optimization Algorithmic Recourse

Open paper AskPDF arXiv

Actionable Recourse provides individuals with actions they can take to change an unfavorable classifier outcome. While useful at the instance level, it is ill-suited for global auditing and bias detection, since aggregating local actions is costly and often inconsistent. Recourse Summaries address this limitation by partitioning the population and assigning one shared action per subgroup, enabling comparison across subgroups. Designing summaries involves a fundamental trade-off between recourse effectiveness and recourse cost, which existing methods do not adequately address. We introduce Summaries of Optimal and Global Actionable Recourse (SOGAR), which formulates recourse summary learning as an optimal decision tree learning problem and finds the Pareto front -- the complete set of solutions where improving one objective necessarily worsens the other. SOGAR enables post-hoc selection of the desired trade-off without retraining. Using shallow axis-parallel decision trees and sparse leaf actions, SOGAR produces stable, low-cost, and effective recourse summaries that outperform existing approaches across effectiveness and cost metrics.

Abstract

Algorithmic Recourse

Explore similar work

Optimal Recourse Summaries via Bi-Objective Decision Tree Learning

DRACULA: Hunting for the Actions Users Want Deep Research Agents to Execute

Algorithmic Recourse Under Competition