Simple Network Graph Comparative Learning

• The paper introduces SNGCL, a novel method leveraging asymmetric metrics and contrastive learning to encode directed relationships in network data.
• The methodology employs dual encoder architectures and binary NCE loss to capture reachability and direction-dependent costs in graph-structured models.
• The approach benefits applications like subgoal discovery, planning, and hierarchical clustering, demonstrating improved performance in asymmetric environments.

Simple Network Graph Comparative Learning (SNGCL) refers to a class of machine learning techniques that leverage network (graph) structure during representation learning using comparative objectives—typically through asymmetric or contrastive distance functions tailored to directed or non-symmetric relationships. SNGCL unifies recent advances in representation learning, asymmetric metrics, contrastive embedding architectures, and downstream planning or clustering tasks on network-structured data. The approach enables the learning of embeddings that explicitly represent directed relationships (such as reachability or containment) in the underlying graph or state space, with applications in areas including reinforcement learning, subgoal discovery, network analysis, and more.

1. Asymmetric Distance Functions in Network Representation

A central mathematical component of SNGCL is the asymmetric distance (or similarity) metric, designed to capture non-symmetric relationships between node pairs in a network. Unlike symmetric metrics (e.g., Euclidean or cosine distances), asymmetric distances satisfy $d(x, y) \neq d(y, x)$, reflecting the directionality or causality present in the network topology or data generation process.

Recent work on probabilistic world modeling formalizes the learned asymmetric similarity $f(x, y)$ as a function of separate encodings for source and target nodes (or states): $f(x, y) = d(\psi(x), \phi(y))$ where typically $\psi$ and $\phi$ are parameterized (and distinct) neural encoders for node $x$ ("target") and node $y$ ("source") respectively (Song, 2024). The specific formulation and choice of $d(\cdot, \cdot)$ may range from asymmetric cosine similarity to asymmetric neural semi-norms such as "WideNorm," which operate on concatenated positive and negative parts of latent differences (Steccanella et al., 2023).

This asymmetry encodes, for instance, the property that in a state-transition graph, the existence of a path from $y$ to $x$ does not imply a path in the opposite direction. Thus, the learned distance provides a geometric abstraction of directed reachability or containment suited for directed graphs and Markov decision processes.

2. Comparative Learning Objectives and Contrastive Losses

SNGCL methods employ comparative or contrastive learning objectives that are explicitly adapted to the asymmetry of the domain. In deep representation learning, a standard technique is contrastive loss (such as InfoNCE), which encourages proximity between positive (related) pairs and repulsion between unpaired negatives.

In the asymmetric setting, the construction of positive and negative pairs is dictated by the network (or state) transitions:

• Positive pairs: represent reachable pairs in the directed network (e.g., consecutive states in a trajectory, or source-to-target pairs along network edges).
• Negative pairs: are drawn independently, often from the marginal distribution of nodes or states (Song, 2024).

A canonical instance is the binary Noise-Contrastive Estimation (NCE) loss: $$J_{BINCE} = \mathbb{E}_{(x^+, y)} [\log \sigma(f(x^+, y))] + K \cdot \mathbb{E}_{y, x^-} [\log (1 - \sigma(f(x^-, y)))]$$ where $x^+$ is a node reachable from $y$ and $x^-$ is a negative sample. The asymmetry is preserved during parameterization and sampling.

Such objectives, in SNGCL, serve to structure the latent representation space in a manner that reflects the "one-way" comparatives inherent in the data distribution or network.

3. Embedding Architectures and Implementation

Key to SNGCL is the design of encoder architectures that can capture and utilize directionality:

• Dual or Split Encoders: Separate encoders $\phi$ and $\psi$ for source and target nodes or states, ensuring that the resulting embedding is not forced to be symmetric. This is essential for the mathematical realization of $d(x, y) \neq d(y, x)$ (Song, 2024).
• Asymmetric Norms: Architectures such as WideNorm employ learned linear maps on positive and negative differences in the latent space, producing asymmetric (semi-)norms that respect triangle inequalities and encode directionality in action or transition costs (Steccanella et al., 2023).
• Graph Structure Conditioning: Embedding network structure via positional, neighborhood, or message-passing operations is common. However, what distinguishes SNGCL is the comparative component that leverages trajectory or pathwise information to encode long-horizon asymmetric relations.

Computationally, these architectures are efficient: computing asymmetric distances reduces to standard neural embedding evaluations and vector operations, and the training scales with the number of sampled contrasts (not the network size).

4. Applications: Subgoal Discovery, Planning, and Network Analysis

SNGCL has demonstrated particular effectiveness in:

• Subgoal Discovery in Reinforcement Learning: By clustering in the asymmetric-distance-based latent space (typically after re-symmetrizing locally from a reference state), critical bottleneck or bridge nodes (such as doorways or passageways in gridworlds) emerge as outliers, corresponding to intuitively meaningful subgoals for hierarchical planning (Song, 2024).
• Goal-Conditioned Planning: The asymmetric metric provides an informative and admissible heuristic for planning, especially in environments with direction-dependent transition costs. Empirical results show superior performance to symmetric metrics in settings where reverse transitions are expensive or impossible (Steccanella et al., 2023).
• Hierarchical Structure Modeling: In tasks such as hierarchical classification on trees or partial orders, asymmetric comparative losses and metrics more faithfully encode the semantic cost or risk associated with moving from node $y$ to node $x$, respecting the hierarchy's structure (Mekala et al., 2018).
• Directed Graph Representation: SNGCL offers a principled method for lifting graph directionality into the learned latent space, enabling algorithms that exploit reachability, minimum action distance, and sequence constraints.

A summary table of core SNGCL operations and their roles:

Operation	Implementation	Functionality
Asymmetric Distance	$f(x, y) = d(\psi(x), \phi(y))$	Encodes directionality, reachability
Contrastive Loss	Binary NCE, InfoNCE with trajectory pairs	Enforces representation of comparative relations
Subgoal Extraction	DBSCAN or density on embedding (after reference-conditioning)	Identifies bottlenecks and usable subgoals
Planning/Control	Action selection minimizing asymmetric distance	Drives agents optimally toward goals in asymmetric spaces

5. Theoretical Properties and Metric Guarantees

SNGCL methods rely on the mathematical properties of the asymmetric metrics utilized:

• Triangle Inequality: Many asymmetric (semi-)norm constructions automatically satisfy a generalized triangle inequality: $d(x, z) \leq d(x, y) + d(y, z)$, even if $d(x, y) \neq d(y, x)$ (Steccanella et al., 2023).
• Directed Reachability and Containment: The metric structure is such that $d(y, x)$ reflects the cost, number of actions, or probability mass of moving from $y$ to $x$, matching application requirements (e.g., $d(x, y) = 0$ iff $x$ is reachable from $y$ at zero cost, which is only true in networks with direct edges).
• Reference Conditioning: Symmetrization conditioned on a reference state is possible, yielding a metric suitable for clustering while retaining global reachability information (Song, 2024).

These theoretical properties are essential for the soundness of downstream algorithms—such as subgoal discovery, hierarchical planning, and error-bounded inference in learned models.

6. Empirical Evaluation and Comparative Performance

Empirical studies illustrate the superiority of SNGCL's asymmetric metrics over symmetric alternatives when the environment exhibits inherent asymmetry:

• In symmetric environments, both SNGCL and symmetric embeddings perform equivalently (e.g., similar MSE to ground-truth distances and success rates in planning).
• In asymmetric environments (e.g., where backward movement is costly), asymmetric embeddings preserve the proper cost structure, yielding low error and much higher planning success rates compared to symmetric metrics, which tend to "collapse" distances and mislead planners (Steccanella et al., 2023).
• Subgoal discovery by clustering in the asymmetric latent space consistently identifies true bottlenecks (such as doors or corridor midpoints) as outliers, matching intuitive decompositions of long-horizon tasks (Song, 2024).

7. Connections, Extensions, and Open Problems

SNGCL is closely related to recent developments in asymmetric metric learning, hierarchical loss functions, and directed-graph neural embeddings. Conceptual links can be drawn to:

• Asymmetric distance codes and related combinatorial optimizations (Gabrys et al., 2015)
• The extension of classic symmetric metrics (e.g., Fubini–Study, Frobenius) to the asymmetric setting on spaces such as Grassmannians (Mandolesi, 2023, Mandolesi, 2022)
• General streaming and query-based algorithms for asymmetric edit and action distances (Saks et al., 2012, Andoni et al., 2010, Li et al., 2021)

Current open problems include scalable adaptation to large and dynamic graphs, characterization of metric properties under varying asymmetries, and the development of tighter theory for subgoal optimality and planning guarantees in highly non-symmetric state spaces.

---

References:

• "Probabilistic World Modeling with Asymmetric Distance Measure" (Song, 2024)
• "Asymmetric Norms to Approximate the Minimum Action Distance" (Steccanella et al., 2023)
• "Bayes-optimal Hierarchical Classification over Asymmetric Tree-Distance Loss" (Mekala et al., 2018)
• "Asymmetric Lee Distance Codes for DNA-Based Storage" (Gabrys et al., 2015)
• "Asymmetric Metrics on the Full Grassmannian of Subspaces of Different Dimensions" (Mandolesi, 2022)
• "Asymmetric Geometry of Total Grassmannians" (Mandolesi, 2023)
• "Space efficient streaming algorithms for the distance to monotonicity and asymmetric edit distance" (Saks et al., 2012)
• "Polylogarithmic Approximation for Edit Distance and the Asymmetric Query Complexity" (Andoni et al., 2010)
• "Lower Bounds and Improved Algorithms for Asymmetric Streaming Edit Distance and Longest Common Subsequence" (Li et al., 2021)