Research Portfolio Optimization

• Research Portfolio Optimization is a binary selection framework that uses semantic analysis of publication content to predict future academic impact.
• It leverages a temporal knowledge graph and NLP-based features like originality and diffusion measures to assess each publication’s potential influence.
• The model maximizes a risk-adjusted score under capacity constraints, outperforming random allocation while addressing prediction uncertainty.

Research portfolio optimization is a forward-looking selection problem over scientific outputs in which scarce attention, funding, or publication capacity is allocated to items expected to generate the greatest future intellectual influence. In the formulation developed by the rXiv Semantic Impact framework, the objective is not citation accumulation but predicted future Semantic Impact (SI), measured through the evolution of a temporal academic knowledge graph built from publication semantics. The resulting optimization is a binary, capacity-constrained, risk-penalized selection model: choose a subset of available publications to maximize predicted future semantic influence while discounting prediction uncertainty (Belikov, 19 Feb 2025).

1. Motivation and conceptual scope

The motivating premise is that citation-based research allocation is distorted by social and institutional biases, including journal prestige, author visibility, and institutional reputation. Under this view, portfolio decisions based purely on citations can misallocate scientific resources, especially when early recognition is weak for work that later becomes scientifically important. The XSI framework therefore replaces citation counts with a knowledge-graph-based signal that asks whether the semantic content of a paper becomes more central, reused, or “popular” in the future fabric of science (Belikov, 19 Feb 2025).

Within this framework, a “portfolio” is not a continuous budget split across topics in the classical finance sense. It is a binary selection problem over available publications in a given period. A funding agency, journal, lab, or researcher chooses a subset of publications—or, by analogy, proposals—to maximize predicted future semantic impact while penalizing predictive uncertainty. The paper is explicit that this differs from standard Markowitz portfolio selection in several ways: there is no covariance matrix across selected items, no risk-free asset, no shorting, and no fractional allocation. The optimization is therefore better described as a budget-constrained binary selection problem with additive utility and additive uncertainty penalty than as a mean-variance quadratic program (Belikov, 19 Feb 2025).

A common misconception is that research portfolio optimization in this setting is merely citation forecasting with different notation. It is not. The target variable is designed to capture future semantic uptake, not contemporaneous popularity. The framework is intended as a complement to citations, not a relabeling of them (Belikov, 19 Feb 2025).

2. Knowledge-graph substrate and Semantic Impact

The framework is built on a temporal academic knowledge graph derived from 324K open-access biomedical publications from bioRxiv, medRxiv, and arXiv q-bio, covering 2003–2025 in the abstract; the main text specifies data as of December 24, 2024. Each publication abstract is parsed by the authors’ NLP pipeline TriEL, which performs relation extraction, co-reference resolution, and entity linking via BERN2 and Entity Fishing. Each publication $\pi$ is represented as a directed entity graph $G_\pi$, and the global graph is

$G = \bigcup_{\pi} G_\pi.$

To construct the reduced temporal graph sequence used for analysis, the framework retains only edges from publications available before time $t$, removes non-informative vertices such as type $T_0$, and reconnects paths through removed vertices. If vertices $a$ and $b$ were connected through a removed vertex $u$, the rewired edge weight is

$$w_{ab}^{-1} = \left(w_{au}^{-1} + w_{ub}^{-1}\right)^{-1}.$$

For practical reasons, the reduced graph is built as an approximate rolling union,

$$G_r(t) = \bigcup_{t-\Delta \le t_\pi \le t} G_{\{\pi\}}, \qquad \Delta = 10 \text{ years}.$$

The central target variable is Semantic Impact. For publication subgraph $G_\pi$0, the raw SI is

$G_\pi$1

and the transformed SI used for modeling is

$G_\pi$2

This definition measures the average proportional increase in the weights of the semantic relations expressed by the paper between publication time $G_\pi$3 and future time $G_\pi$4. The intended edge cases are explicit: work that is semantically popular at $G_\pi$5 but not at $G_\pi$6 should have low SI, while work whose semantics become popular only later should have high SI.

A closely related feature is originality, defined as the inverse average edge weight at publication time:

$G_\pi$7

Lower prior prevalence of a paper’s semantic edges therefore implies higher originality (Belikov, 19 Feb 2025).

3. Predictive modeling in the XSI framework

The predictive task is to estimate, at publication time, a paper’s future semantic impact over horizons

$G_\pi$8

The framework uses only semantic and graph-structural information available at publication time. Features include originality; degree-related summaries; diffusion-based measures over up to four steps; community-related descriptors based on label propagation and KS-type distributional comparisons; vertex novelty indicators; and same-community flags. The paper describes these as operational manifestations of “semantic trajectories” in the evolving graph (Belikov, 19 Feb 2025).

Two model classes are evaluated: Lasso linear regression and Histogram-based Gradient Boosting Regression Trees (HBR) from scikit-learn. Training uses temporal rolling windows, and the paper explicitly notes that both SI and features are non-stationary; it therefore normalizes SI using a 180-day moving average and uses transformed targets and scaled features in the final setup. The framework also models predictive uncertainty through the relative error surrogate

$G_\pi$9

which later becomes the portfolio “risk” term (Belikov, 19 Feb 2025).

Empirically, HBR outperforms Lasso. The appendix reports out-of-sample $G = \bigcup_{\pi} G_\pi.$0 for 6-month SI prediction and $G = \bigcup_{\pi} G_\pi.$1 for 36-month SI prediction; the abstract highlights $G = \bigcup_{\pi} G_\pi.$2 three years in advance. For the SI prediction error model, the reported $G = \bigcup_{\pi} G_\pi.$3 values are $G = \bigcup_{\pi} G_\pi.$4 at 6 months and $G = \bigcup_{\pi} G_\pi.$5 at 36 months (Belikov, 19 Feb 2025).

Feature-importance analysis shows that originality is the dominant predictive signal, with diffusion features second. Citation correlation is positive but modest: the paper reports typical Spearman correlations in the range $G = \bigcup_{\pi} G_\pi.$6 to $G = \bigcup_{\pi} G_\pi.$7, with examples around $G = \bigcup_{\pi} G_\pi.$8 for PLOS One, $G = \bigcup_{\pi} G_\pi.$9 for Nature Communications, and $t$0 for eLife. This supports the claim that SI is related to, but not reducible to, citation counts (Belikov, 19 Feb 2025).

4. Optimization model and solver

The portfolio optimization layer is introduced explicitly as an analogue of financial portfolio selection, but the mathematical structure is different. The decision variables are binary:

$t$1

where $t$2 is the set of publications available in the relevant period, and $t$3 means publication $t$4 is selected. The exact optimization problem is

$t$5

subject to

$t$6

$t$7

Here $t$8 is the predicted semantic impact, $t$9 is predicted uncertainty, $T_0$0 is the maximum fraction of publications that may be selected, and $T_0$1 is a risk-aversion parameter (Belikov, 19 Feb 2025).

This formulation is additive: it maximizes a risk-adjusted score $T_0$2 across selected items, under a capacity constraint. Because there is no covariance structure, no diversification term, and no explicit dependence penalty among selected publications, the model is mathematically closer to a cardinality-constrained knapsack-like integer linear program than to classical Markowitz mean-variance optimization. The paper states that it is solved using Integer Linear Programming, specifically Google OR-Tools and CP-SAT (Belikov, 19 Feb 2025).

A common misunderstanding is that the knowledge-graph setting automatically induces topic diversification in the optimization itself. It does not. The graph is used to construct SI and predictive features; the optimization objective does not directly enforce topic diversity, community coverage, novelty quotas, or redundancy penalties among selected papers (Belikov, 19 Feb 2025).

5. Empirical behavior and intended use cases

The optimization experiments are temporal and out of sample. Models for $T_0$3 and $T_0$4 are trained on historical windows, then used to select portfolios in later periods, and finally evaluated by realized future SI. The paper presents portfolio experiments for $T_0$5 and $T_0$6 months and highlights selection fractions of 5%, 10%, and 20%. The benchmark is random selection of the same fraction of available publications. Because SI is normalized, random selection yields realized performance approximately equal to unity (Belikov, 19 Feb 2025).

The principal empirical result is that the optimization framework systematically outperforms random allocation. Predicted portfolios outperform realized portfolios, as expected under forecast error, but realized optimized portfolios remain clearly above the random baseline. The gains appear strongest when selecting smaller fractions—first 5%, then 10%, then 20%—which is consistent with a ranking model whose top-scored items are most enriched for future semantic impact (Belikov, 19 Feb 2025).

The intended applications are broad. The paper suggests that funding agencies, publishers, institutions, labs, and individual researchers can use XSI-based portfolio optimization to guide selection decisions. A funder could score proposals or outputs by predicted future SI and choose a subset under a capacity constraint; a journal could use SI-based ranking as a complement to peer review; a lab could allocate attention toward research lines expected to have high future semantic influence (Belikov, 19 Feb 2025).

The paper also reports a notable disturbance around early 2020 due to COVID-19, which changed SI distributions and temporarily worsened predictive performance. This is important because it shows that external shocks can alter graph dynamics and thereby the forecasting environment on which portfolio selection depends (Belikov, 19 Feb 2025).

6. Limitations, controversies, and open directions

The framework has several explicit limitations. First, it is empirically restricted to biomedical preprints and q-bio arXiv papers; generalizability beyond that domain is asserted in principle but not demonstrated. Second, the approach is strongly dependent on knowledge graph quality. Entity-linking errors can directly distort both targets and features; the paper gives the example of semaglutide being incorrectly mapped to semaphorin 5B. Third, the graph is built only from abstracts, not full texts, which necessarily limits semantic coverage (Belikov, 19 Feb 2025).

A deeper conceptual limitation is that SI measures what the paper calls the declarative quality of research rather than the correctness or execution quality of the work. A paper can receive high SI if its semantic claims later become influential, even if later work supplies the real proof or validation. Relatedly, the framework has a first-mover advantage: earlier papers that introduce a claim can score higher than later papers that substantiate it. The paper also notes that publication performance is initially treated as independent, so the optimization omits a core element of classical portfolio theory: dependence across selected items (Belikov, 19 Feb 2025).

Another important misconception is that semantic-impact optimization is automatically bias-free. The paper does not make that claim. Its argument is narrower: SI-based optimization may mitigate some citation-related biases because its biases are different from those of prestige- and visibility-driven citation systems. A plausible concern is that if SI becomes an explicit target for funding or publishing, researchers may adapt behavior to maximize semantic uptake rather than truth, rigor, or broader scientific value. The paper hints at this indirectly by noting that SI does not measure correctness and by proposing future extensions to estimate whether claims “hold up over time” (Belikov, 19 Feb 2025).

The main methodological open directions are therefore clear: incorporate inter-item dependence and diversification, broaden the corpus beyond abstracts and beyond biomedicine, improve entity linking for emerging concepts, and align optimization targets more closely with long-run epistemic quality rather than semantic diffusion alone. In its current form, research portfolio optimization in XSI is best understood as a technically explicit, knowledge-graph-driven, uncertainty-aware binary selection framework for choosing scientific items with high predicted future semantic influence (Belikov, 19 Feb 2025).