Diversity-Driven RandomWalks for News Recommenders

• Diversity-Driven RandomWalks (D-RDW) are methods that integrate random-walk candidate generation with explicit target distributions to enforce editorial diversity in recommendations.
• The approach leverages multi-hop random walks on a user–article bipartite graph and constraint-based sampling to attain preset proportions of attributes like sentiment and political party mentions.
• Empirical evaluations indicate that D-RDW achieves a balanced trade-off between accuracy, diversity, and computational efficiency, offering a cost-effective and transparent recommender solution.

Searching arXiv for the named method and closely related random-walk diversification papers. arXiv search query: "D-RDW Diversity-Driven Random Walks news recommender systems random walks with erasure directed random walks negative reinforcement" Diversity-Driven RandomWalks (D-RDW) denotes a family of recommendation and ranking methods that use modified random-walk dynamics or random-walk-based selection procedures to improve diversity without discarding graph-based relevance signals. In the most explicit formulation, D-RDW is a lightweight algorithm and re-ranking technique for news recommendation that combines random walks on a user–article graph with customizable Normative Target Distributions (NTDs) over article properties such as sentiment and political party mentions, thereby treating diversity as a controllable in-processing objective rather than only a post-hoc constraint (Li et al., 18 Aug 2025). In a broader historical sense, D-RDW also describes a line of work in which diversity is induced by directing similarity flow toward low-degree users, by propagating negative reinforcement in Personalized PageRank, or by selectively erasing walk mass and reinjecting it to the source, all with the shared aim of counteracting popularity concentration, redundancy, or ideological homophily (Liu et al., 2012, Badrinath et al., 2012, Paudel et al., 2021).

1. Conceptual scope and historical lineage

The explicit term “D-RDW” appears in “D-RDW: Diversity-Driven Random Walks for News Recommender Systems” (Li et al., 18 Aug 2025), where it names a news recommender that integrates random-walk candidate generation with constraint-based sampling over interpretable article properties. The paper frames the method as a societal recommender and links it to normative and democratic news recommender systems, emphasizing that diversity in political news is not merely a secondary utility metric but a mechanism for operationalizing editorial and societal values (Li et al., 18 Aug 2025).

Earlier work established closely related mechanisms without using the D-RDW label. “Solving the accuracy-diversity dilemma via directed random walks” introduced the NCF algorithm, a collaborative filtering method based on directed random walks on a user–object bipartite graph, where the key intervention is to use similarity from neighbors to the target user so that low-degree users exert greater influence and niche items are more likely to surface (Liu et al., 2012). The paper’s central theoretical identity,

$\frac{s_{ij}}{s_{ji}}=\frac{k_{u_i}}{k_{u_j}},$

shows that directed similarity is reverse to the initial node’s degree and provides a formal explanation for why changing the direction of random-walk similarity can improve diversity while maintaining or improving accuracy (Liu et al., 2012).

A second branch is “Diversity in Ranking using Negative Reinforcement,” which proposed Negative Reinforcement Ranking (NR2), an iterative top-$k$ graph ranking method based on repeated Personalized PageRank solves with negative prior mass assigned to already selected nodes (Badrinath et al., 2012). NR2 is not a one-shot random walk but a greedy diversity-driven procedure in which anti-redundancy signals are propagated over the graph through the restart vector. A third related branch is “Random Walks with Erasure: Diversifying Personalized Recommendations on Social and Information Networks,” which introduced Random Walk with Erasure (RWE), where source–destination-specific erasure probabilities suppress some walk mass and reinject it to the source, shifting recommendations toward ideologically diverse or long-tail items (Paudel et al., 2021).

These papers do not define a single unified D-RDW formalism. This suggests that the term functions both as the proper name of the 2025 news method and as a useful umbrella for diversity-aware random-walk recommenders that modify walk direction, restart behavior, endpoint retention, or candidate selection in order to alter the composition of the final ranking.

2. Core formulation in news recommendation

In the 2025 formulation, D-RDW operates on a bipartite graph of user nodes and item or article nodes, with edges representing clicks (Li et al., 18 Aug 2025). The method is organized into three stages: pre-processing, in-processing, and post-processing. Pre-processing constructs and augments the graph; in-processing performs random walks and distribution-constrained sampling; post-processing ranks the selected items and may optionally apply graph coloring for more homogeneous category presentation (Li et al., 18 Aug 2025).

The random-walk stage begins from the target user and uses 3 hops on the user–article graph. Because the graph is bipartite, odd hop counts terminate on articles, and the pseudocode increases hop count by $2$ when necessary, yielding the sequence

$3,5,7,\dots,\texttt{maxHops}.$

The paper states that candidate items can be sorted by random walk transition probability or prediction score, but it does not provide an explicit transition-matrix equation, personalized PageRank equation, teleportation term, or restart law (Li et al., 18 Aug 2025). Accordingly, the walk is specified procedurally rather than through a closed-form Markov formulation.

The distinctive mechanism is the Normative Target Distribution (NTD). After walk-based candidate generation and filtering of items already in the user’s history, D-RDW selects a subset whose composition matches specified target proportions over interpretable article properties. The paper presents this as a binary selection problem with a binary vector $X$, where $x_i\in\{0,1\}$ indicates whether candidate item $i$ is chosen. In the illustrative example,

$$\begin{pmatrix} I^\intercal \ P^\intercal \ S^\intercal \end{pmatrix} X= \begin{pmatrix} 20\ 15\ 10 \end{pmatrix},$$

so the list must contain exactly $20$ items, $15$ satisfying the political criterion, and $k$0 satisfying the sentiment criterion (Li et al., 18 Aug 2025). The sampling objective is

$k$1

where $k$2 can be random walk transition probability or another numeric item feature such as recency (Li et al., 18 Aug 2025).

If the constraints are infeasible, D-RDW reduces the target size by one until it finds the maximum-sized subset satisfying both the NTD ratios and the adjusted target size (Li et al., 18 Aug 2025). If there are still fewer than targetSize items after the walk-and-sample loop, the system fills the remaining positions with randomly selected articles and then ranks the final recommendations (Li et al., 18 Aug 2025).

This architecture differs from standard re-ranking because diversity is imposed during recommendation generation rather than only after a candidate list is fixed. It also differs from earlier walk diversification methods because the diversity criterion is encoded as an explicit distribution over article attributes rather than only through altered transition dynamics.

3. Normative target distributions and controllable diversity

A defining property of D-RDW is that diversity is specified through editorially interpretable distributions over article properties. In the main experiment, the target dimensions are sentiment and political party mentions (Li et al., 18 Aug 2025). Sentiment is discretized into four buckets with target proportions:

• $k$3: $k$4
• $k$5: $k$6
• $k$7: $k$8
• $k$9: $2$0

Political party mentions are divided into five buckets:

1. government parties (including supporting parties): $2$1
2. opposition parties: $2$2
3. both government and opposition: $2$3
4. independent/foreign parties: $2$4
5. no political parties: $2$5 (Li et al., 18 Aug 2025)

The paper treats these distributions as one operationalization of a deliberative recommender and explicitly associates party mentions with viewpoints and sentiment with emotion (Li et al., 18 Aug 2025). Editors can also choose which dimensions matter, what proportion of the list each bucket should occupy, how large the recommendation list should be, and what objective should be optimized among feasible items (Li et al., 18 Aug 2025). The paper discusses category as another interpretable property, although it is not included in the main NTD (Li et al., 18 Aug 2025).

This degree of controllability is central to the method’s normative interpretation. Rather than leaving diversity as an emergent property of graph structure, D-RDW makes recommendation composition directly parameterizable in human-readable terms such as “15% opposition-party articles” or “20% strongly negative sentiment” (Li et al., 18 Aug 2025). A plausible implication is that D-RDW shifts part of recommender design from latent optimization toward explicit policy specification.

The constraint-based sampling mechanism also sharply distinguishes D-RDW from diversity-aware methods that encode diversity indirectly. In NCF, diversity emerges because low-degree users receive more influence under directed similarity (Liu et al., 2012). In NR2, diversity emerges because negative prior mass demotes nodes near already selected ones (Badrinath et al., 2012). In RWE, diversity emerges because source–destination-specific erasure probabilities suppress ideologically similar or high-degree endpoints (Paudel et al., 2021). D-RDW instead exposes target diversity as a user- or editor-defined output distribution (Li et al., 18 Aug 2025).

4. Relation to precursor random-walk diversification methods

The D-RDW name is most precise for the 2025 news recommender, but its design belongs to a broader methodological lineage of diversity-aware random walks.

Directed random walks and NCF

In the NCF algorithm, the system is represented as a user–object bipartite graph

$2$6

with adjacency matrix

$2$7

where $2$8 if object $2$9 is collected by user $3,5,7,\dots,\texttt{maxHops}.$0, and $3,5,7,\dots,\texttt{maxHops}.$1 otherwise (Liu et al., 2012). The directed similarity from user $3,5,7,\dots,\texttt{maxHops}.$2 to user $3,5,7,\dots,\texttt{maxHops}.$3 is

$3,5,7,\dots,\texttt{maxHops}.$4

and the prediction rule is

$3,5,7,\dots,\texttt{maxHops}.$5

Because

$3,5,7,\dots,\texttt{maxHops}.$6

the similarity is inversely related to the degree of the source user. Reorienting similarity from neighbors to the target user therefore amplifies low-degree users, which the paper argues helps reveal niche objects and solve the accuracy-diversity dilemma (Liu et al., 2012).

Negative reinforcement in Personalized PageRank

NR2 uses a row-normalized similarity matrix $3,5,7,\dots,\texttt{maxHops}.$7, a preference vector $3,5,7,\dots,\texttt{maxHops}.$8, and damping factor $3,5,7,\dots,\texttt{maxHops}.$9, with Personalized PageRank defined by

$X$0

or equivalently

$X$1

After each selection, the prior is modified so that already selected nodes receive negative mass, unselected nodes receive the remaining positive mass, and an absorbing node $X$2 receives mass $X$3: $X$4 This propagates negative influence to nearby vertices and causes the next argmax to come from a different central region (Badrinath et al., 2012). Relative to D-RDW, NR2 is diversity-driven in mechanism but greedy and iterative, not a unified recommender over the full candidate space.

Random Walk with Erasure

RWE starts from a finite odd-length walk on a user–item graph with transition matrix

$X$5

and then applies a source–destination-specific erasure matrix

$X$6

If walk mass $X$7 reaches destination $X$8 from source $X$9, the method erases $x_i\in\{0,1\}$0, retains $x_i\in\{0,1\}$1, and returns the erased mass to the source for the next iteration (Paudel et al., 2021). For long-tail diversification,

$x_i\in\{0,1\}$2

so high-degree items receive higher erasure. For ideological bridging,

$x_i\in\{0,1\}$3

with $x_i\in\{0,1\}$4 (Paudel et al., 2021). The method thus implements diversity by selective endpoint suppression and source reinjection.

Across these variants, the common principle is that a random walk’s default bias toward popular, central, or nearby regions is modified so that underrepresented regions of the graph receive more probability mass or more final-list representation.

5. Experimental evaluation and empirical behavior

The most detailed D-RDW evaluation concerns news recommendation on EB-NeRD small, drawn from the RecSys 2024 Challenge benchmark (Li et al., 18 Aug 2025). The reported test setup includes 15,339 users, an article pool of 4,400+ articles, a median of 11 impression items per user session, and a median of 1 clicked item per session (Li et al., 18 Aug 2025). Because diversity metrics require larger lists, the evaluation ranks over the entire test pool to generate top-20 recommendation lists rather than limiting ranking to the original impressions (Li et al., 18 Aug 2025).

The baselines comprise neural news recommenders (LSTUR, NPA, NRMS), re-ranking baselines (G-KL, PM-2, MMR), random-walk baselines (3, RWE-D), and a Random baseline (Li et al., 18 Aug 2025). Accuracy is measured by AUC, while diversity is evaluated using RADio metrics, Gini, ILD, and computational cost metrics (Li et al., 18 Aug 2025).

Accuracy and cost

Reported AUC values are:

• 3: 0.565
• LSTUR: 0.564
• D-RDW: 0.554
• RWE-D: 0.554
• NPA: 0.554
• NRMS: 0.549
• Random: 0.500 (Li et al., 18 Aug 2025)

These results show that D-RDW is competitive but not the strongest pure-accuracy model. The paper’s claim is therefore not that D-RDW dominates AUC, but that it offers a better overall compromise between accuracy, diversity, and cost (Li et al., 18 Aug 2025).

Training and recommendation cost were reported in watt-seconds:

Method	Train Cost	Recommendation Cost
LSTUR	351.12	69.84
NPA	523.05	51.25
NRMS	326.55	63.23
D-RDW	2.81	11.76
3	2.75	1.27
RWE-D	3.15	0.27

These numbers support the claim that random-walk methods are about one order of magnitude cheaper than the neural baselines in this setting (Li et al., 18 Aug 2025). D-RDW is more expensive than simpler random-walk baselines because the NTD sampling step adds overhead, but it remains much cheaper than the neural models (Li et al., 18 Aug 2025).

Diversity results

On the norm-relevant dimensions, D-RDW exactly matches the NTV targets on both Gini and ILD:

• Sentiment Gini = 0.133
• Party Gini = 0.250
• Sentiment ILD = 0.779
• Party ILD = 0.789 (Li et al., 18 Aug 2025)

The paper also reports strong RADio values for D-RDW:

• Activation: 0.374
• Category Calibration: 0.407
• Complexity Calibration: 0.229
• Fragmentation: 0.394
• Alternative Voices: 0.107
• Representation: 0.556 (Li et al., 18 Aug 2025)

The interpretation given in the paper is that D-RDW is the best overall compromise across normative diversity, traditional diversity, computational cost, and AUC, with “consistent scores in the top 3 across all four dimensions” (Li et al., 18 Aug 2025). At the same time, the results do not imply that D-RDW is the best on every metric. For example, 3 has slightly higher AUC and lower recommendation cost, and Random has the strongest category ILD and category Gini, which is unsurprising because category is not part of D-RDW’s NTD (Li et al., 18 Aug 2025).

A related point is that the underlying corpus distribution is highly imbalanced. For party mentions, the EB-NeRD distribution is $x_i\in\{0,1\}$5 government/supporters, $x_i\in\{0,1\}$6 opposition, $x_i\in\{0,1\}$7 both, $x_i\in\{0,1\}$8 independent/foreign, and $x_i\in\{0,1\}$9 no political entities; sentiment is $i$0 negative, $i$1 somewhat negative, $i$2 somewhat positive, and $i$3 positive (Li et al., 18 Aug 2025). Any method that matches the target distribution therefore diverges substantially from the raw corpus. This suggests that D-RDW’s diversity gains depend not only on walk-based exploration but on the ability of its sampling stage to impose a composition that the data distribution would not naturally produce.

6. Interpretation, limitations, and adjacent theory

D-RDW occupies a distinctive position among random-walk recommender methods because it combines graph-based recommendation with explicit distributional control. This makes it transparent and controllable in a way that is atypical of latent-factor or neural ranking models (Li et al., 18 Aug 2025). It also differs from most prior diversity-aware random-walk methods by exposing what counts as diversity in externally specified, interpretable dimensions.

Several limitations are explicit in the literature. The D-RDW paper does not provide formal transition equations, normalization details, teleportation or restart mechanics, or a complete probabilistic derivation of the random walk itself (Li et al., 18 Aug 2025). This limits reproducibility from the text alone. The same paper also acknowledges that diversity outcomes depend strongly on the dataset distribution, that the chosen NTD is normative and contestable, that personalization of diversity is limited because the same NTD is broadly applied, and that offline gains do not establish user acceptance or engagement effects (Li et al., 18 Aug 2025).

The earlier walk-diversification papers expose complementary limitations. NCF depends on the degree structure of the user–item graph and is especially compelling when many users have low degree (Liu et al., 2012). NR2 is iterative and optimized for top-$i$4 selection, not full-list ranking over a large candidate space (Badrinath et al., 2012). RWE offers an expressive source–destination-specific erasure policy but does not provide a full pseudocode specification, a closed-form converged score equation, or explicit complexity analysis (Paudel et al., 2021).

A broader theoretical perspective on random-walk diversity comes from “Revisiting Random Walks for Learning on Graphs,” which studies Random Walk Neural Networks (RWNNs) rather than recommendation directly (Kim et al., 2024). That paper does not define D-RDW, but it argues that random-walk learning can be understood in terms of coverage, cover time, and probabilistic under-reaching, with non-backtracking, degree-biased walks, and restart mechanisms improving exploration efficiency (Kim et al., 2024). This suggests a theoretical lens for D-RDW-like systems: diversity can be interpreted not only as distributional balance in the output list but also as non-redundant exploration of graph regions that standard walks would under-sample.

A further adjacent development is “Information Geometry of Absorbing Markov-Chain and Discriminative Random Walks,” which analyzes Discriminative Random Walks (DRWs) through class-specific hitting-time laws, Fisher information, and sensitivity geometry (Kimura, 9 Feb 2026). That work is not a D-RDW method, but it provides a mathematical framework for reasoning about identifiable perturbation directions and nonredundant node selection. A plausible implication is that future D-RDW variants could combine output diversity with geometrically diverse seed, node, or edge perturbations (Kimura, 9 Feb 2026).

Taken together, the literature shows that D-RDW can refer narrowly to a 2025 news recommender with NTD-constrained random-walk sampling, or more broadly to a class of methods that alter random-walk recommendation so that diversity becomes a first-class design objective rather than a side effect of graph proximity (Li et al., 18 Aug 2025, Liu et al., 2012, Badrinath et al., 2012, Paudel et al., 2021). In the explicit news setting, its defining contribution is the insertion of interpretable, editor-controllable target distributions into the recommendation pipeline. In the broader methodological lineage, its significance lies in demonstrating that changes to random-walk direction, restart mass, endpoint retention, or constrained selection can redistribute recommendation exposure away from popularity-dominated or homogeneous regions of the graph.