NeuroRecon: Constraint-Aware Neuro-Symbolic Learning for Enterprise Financial Reconciliation
Chandrakanth Thadkapally
Manuscript complete. Not submitted; no venue or submission date set.
Enterprise reconciliation requires more than recognising that two records look alike: an accepted match must also satisfy monetary conservation, event ordering, double-entry balance, link cardinality and organisation-specific rules. NeuroRecon is a constraint-aware neuro-symbolic framework that learns transaction similarity and financial validity together rather than sequentially. Heterogeneous ledger, bank, invoice, fee and adjustment records are represented as a typed transaction graph; a graph encoder proposes candidate groups while differentiable penalties turn accounting and business rules into training signal; constrained decoding emits one-to-one, one-to-many, many-to-one and many-to-many reconciliations with per-constraint evidence trails for analyst review. On a controlled benchmark following the public BenchRec schema, and against five matching strategies under identifier corruption, missing records, amount noise, aggregation and reversals, NeuroRecon reaches 94.6% F1 with a 0.6% constraint-violation rate and the best calibration in the comparison. At 75% identifier corruption it retains 80.6% F1, 25.9 points above the pairwise neural model. Ablation shows the amount, balance, cardinality, temporal and business constraints contribute complementary rather than redundant gains. The finding is that financial structure belongs in training, not in an inference-time rule filter.
Keywordsfinancial reconciliation · neuro-symbolic learning · graph neural networks · accounting constraints · transaction matching · explainable AI
- 01Learns similarity and financial validity jointly, so constraints shape representation learning and candidate ranking instead of only rejecting outputs after probability mass has already been allocated.
- 02Formalises five constraint families — amount conservation, temporal ordering, double-entry balance, cardinality and organisation-specific business rules — as differentiable penalties with individually adjustable weights.
- 03Uses constrained decoding to emit 1:1, 1:N, N:1 and N:M groups with a per-constraint evidence trail, so an analyst sees which family was violated rather than an undifferentiated rejection.
- 04Isolates the effect of the constraints with a robustness protocol that corrupts 0–75% of reference identifiers, testing whether a model learned keys or learned financial structure.
- Ledger, bank, invoice, fee and adjustment records become typed nodes in a transaction graph; a graph encoder proposes candidate groups rather than candidate pairs.
- Each constraint family is expressed as a differentiable penalty on the group, so violation is a gradient during training rather than a filter at inference.
- A joint objective balances match quality against constraint residual; hyperparameters were selected on validation F1 subject to a 1% constraint-violation ceiling.
- Constrained decoding turns scored candidates into a consistent set of groups, with an abstention path that routes low-evidence groups to analysts instead of forcing a match.
| Model | Prec. % | Recall % | F1 % | False recon. % | CVR % | ECE |
|---|---|---|---|---|---|---|
| Rule-based | 93.8 | 68.4 | 79.1 | 2.1 | 3.9 | 0.091 |
| Pairwise neural | 87.5 | 85.1 | 86.3 | 6.8 | 8.7 | 0.074 |
| Graph neural | 90.1 | 88.3 | 89.2 | 4.9 | 5.6 | 0.058 |
| Neural + post-filter | 95.0 | 85.7 | 90.1 | 1.4 | 0.8 | 0.046 |
| NeuroRecon | 95.4 | 93.8 | 94.6 | 1.1 | 0.6 | 0.021 |
CVR — constraint-violation rate, the share of accepted groups breaking at least one active rule. Lower is better for false reconciliation, CVR and ECE.
| Configuration | F1 % | Δ vs. full |
|---|---|---|
| NeuroRecon (all constraints) | 94.6 | — |
| − amount conservation | 91.7 | −2.9 |
| − cardinality | 92.1 | −2.5 |
| − double-entry balance | 92.6 | −2.0 |
Temporal and business-rule removals cost less individually but consistently. No single constraint accounts for the gain.
- Post-filtering and joint training are not equivalent. A filter can reject a known-bad group but cannot rescue a valid group the neural model ranked low; joint training moves the ranking itself, which is why F1 rises rather than only precision.
- Identifier corruption is the diagnostic. As reliable keys disappear the pairwise model falls from 86.3% to 54.7% F1, while NeuroRecon holds 80.6% — evidence it learned financial structure rather than key matching.
- Residual violations per 1,000 accepted groups fall from 3.8 to 0.9 for amount and 2.7 to 0.7 for balance, and what remains sits at rounding boundaries or ambiguous reversals — the cases that genuinely warrant a human.
- Calibration improves alongside validity: ECE 0.021 against 0.074 for the pairwise model, which matters because reconciliation teams auto-accept above a confidence threshold and that threshold has to mean something.
- The ablation shows complementary contributions across all five families, which argues for tunable per-constraint weights rather than a single combined rule score.
- The quantitative results come from a controlled BenchRec-compatible snapshot. They are not a rerun of the official dataset and not a live enterprise trial, and should not be read as either.
- An incomplete constraint formulation, or one weighted too heavily, suppresses genuine exceptions — the failure mode is silent, because a suppressed exception looks like a clean match.
- Foreign currency, tax and late adjustment handling require jurisdiction- and organisation-specific allowances that this formulation does not attempt to generalise.
- Candidate blocking bounds achievable recall, so a true group that was never blocked in cannot be recovered downstream.