Pith. sign in

REVIEW 3 major objections 6 minor 4 references

Building a Recommendation System Using Amazon Product Co-Purchasing Network

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A modified GraphSAGE model predicts co-purchasing links better than a Random Forest baseline and random choice, offering a scalable, inductive way to rank new product recommendations.

desk verdict Despite competent network analysis, the paper never tests its central claim of recommending isolated new products, because the evaluation only uses connected nodes and the proposed fix is undefined for degree-0 nodes. read the letter →

arxiv 2506.02482 v1 pith:U6SGN4IE submitted 2025-06-03 cs.SI

Minghao Liu , Catherine Zhao , Nathan Zhou This is my paper
classification cs.SI
keywords co-purchasingnetworklinkpredictionGraphSAGEinductiverecommendationnewitemproblemembeddingscategorysimilarity
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to show that recommending new products can be framed as link prediction on a co-purchase graph, and that a modified GraphSAGE model does this better than a Random Forest baseline. The authors build a graph where products are nodes and co-purchase relations are edges, train on pairs of one-degree nodes and their neighbors, and measure top-k accuracy. If correct, the approach would give e-commerce platforms a scalable, inductive way to rank candidate new products as they are added, using product features and local graph structure rather than past user behavior.

What carries the argument

The central object is a modified GraphSAGE model for link prediction. Standard GraphSAGE aggregates features from a node's neighbors; the modification instead treats the combined features of a node pair as the embedding, samples only one-hop neighbors of the target node, and uses those as a proxy for aggregation. The one-hop rule encodes the assumption that only direct co-purchase relationships, not transitive ones, are relevant for future co-purchases.

What would settle it

Take a product with no co-purchase edges, run the modified GraphSAGE link predictor, and compare its top-k accuracy against a random baseline; if accuracy does not exceed the baseline, the claimed ability to recommend truly new products fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that a modified GraphSAGE model for link prediction can outperform a Random Forest baseline and random choice at ranking co-purchasing links in an e-commerce product graph, and that because the model is inductive it can generalize to products not seen during training. The modification consists of treating the feature vector of a node pair as the embedding, sampling one-hop neighbors of the target node as a proxy for neighborhood aggregation, and deliberately avoiding two-hop sampling because co-purchase is not transitive. The authors construct a co-purchase graph from the dataset, use one-degree nodes as ground truth for link prediction, and evaluate with top-k accuracy on a breadth-first-sampled subgraph of the largest connected component.

Load-bearing premise

The technique for isolated nodes assumes a new product has neighbors to sample from, but a truly new product has none, and the experiments never test that case.

Editorial extensions

If this is right

  • If the approach is correct, new products can be ranked against existing items by predicted co-purchase probability, narrowing the candidate list for manual review.
  • The one-hop sampling assumption implies that co-purchase is a first-order relationship, so indirect connections should not be used for prediction.
  • The model can be updated in real time as new purchase edges appear, without full retraining.
  • The ablation results indicate that category similarity, degree, and clustering coefficient carry most of the predictive signal, while title embeddings contribute little.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the evaluation subgraph is sampled via BFS from the largest connected component, every test node has at least one neighbor; a truly cold-start product with degree zero would receive no neighborhood signal, so the 'new product' claim is as yet unsupported.
  • The method is essentially a pairwise classifier over handcrafted and structural features; the GraphSAGE name refers to the one-hop neighborhood sampling, not to a learned aggregation over multi-hop structure.
  • A natural testable extension is to use temporally split data, where edges after a cutoff date are predicted from earlier edges, to simulate true new products and verify the inductive claim.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes a GraphSAGE-based inductive link prediction method for recommending newly listed products in the Amazon co-purchasing network. The authors build an undirected graph from the SNAP Amazon metadata, extract the largest connected component, generate node features (title embeddings, group one-hots, category vectors, degree, clustering coefficient), train a baseline Random Forest and a modified GraphSAGE model on positive/negative pairs involving degree-1 nodes, and evaluate top-k accuracy on a BFS-sampled connected subgraph. The abstract claims the method outperforms baselines and offers a solution for recommending new, isolated products.

Significance. If the central claim were supported, the approach would offer a scalable, inductive link-prediction method for new-item recommendation in e-commerce, which is a practically relevant problem. The manuscript's strengths are its detailed data-preprocessing pipeline, an ablation study, and a promise of code availability. However, the evaluation does not test the isolated-node regime that is the stated motivation, and the proposed neighbor-sampling mechanism is undefined for degree-0 nodes. The significance as claimed is therefore not established by the experiments; the results support only a weak claim about connected nodes in the same graph component.

major comments (3)
  1. [Sections 2.2 and 4.3] The evaluation never tests the stated problem. Section 2.2 explicitly discards all 159,575 isolated nodes by extracting the largest connected component, and Section 4.3 samples test nodes via BFS from that component, so every test node has degree at least 1. The reported top-k results (e.g., top-5 accuracy 0.0187 for GraphSAGE vs. 0.0125 for Random Forest in Figure 4) therefore support only connected-node predictions, not the abstract's claim of recommending newly listed, isolated products. This is a load-bearing mismatch between the problem statement and the experimental design.
  2. [Section 3.5.2] The proposed modification for isolated nodes is undefined. The text states, "we sample from the target node's neighbors and use these as a proxy for aggregation," but an isolated node has no neighbors by definition, so this operation cannot be executed for the nodes the paper claims to handle. Section 3.3 trains on 1-degree nodes as surrogates, yet 1-degree nodes have a neighbor and are not isolated; no argument establishes that the two regimes are equivalent, and the connected test set cannot verify it.
  3. [Section 4.3] The absolute performance is close to random. At top-5, the random baseline is 0.0063, Random Forest is 0.0125, and Modified GraphSAGE is 0.0187 (Figure 4). Even if restricted to connected nodes, these numbers represent only about 1-2 percent correct recommendations, and no confidence intervals or significance tests are reported, so the claim of "outperforming baseline algorithms" is statistically unsubstantiated.
minor comments (6)
  1. [Section 2.2] "Associativity coefficient" should be "assortativity coefficient."
  2. [Figure 2 caption] "Distirbution" should be "Distribution."
  3. [Section 4.3] "reconmendation" should be "recommendation," and the phrase "in almost all k's" is awkward; use "values of k."
  4. [Resources] "ScentenceTransformer" should be "SentenceTransformer."
  5. [Abstract and full text] The GitHub URL in the full text contains spaces ("final-project-l-minghao z-catherine z-nathan.git") and is not a single clean link; the abstract version without spaces should be used consistently.
  6. [Section 4.4] The discussion claims the model generalizes to unseen nodes without retraining, but no temporal or held-out-node experiment is described; the BFS subgraph is sampled from the same LCC used in training, so the meaning of "unseen" is not clearly established.

Circularity Check

2 steps flagged · score 6.0 of 10

Isolated-node recommendation claim reduces to connected-node link prediction: training and test sets exclude all isolated nodes, and the proposed neighbor-sampling fix is undefined for degree-0 nodes.

  1. renaming known result [Section 2.2, Section 3.3, Section 4.3 and Section 4.4]
    "To reduce training time and avoid issues with isolated nodes lacking network structure and ground truth for link prediction, we extracted the largest connected component (LCC) from the dataset. ... we test them on a subgraph sampled as follows: We randomly select a start node from the largest connected component, then perform a breadth-first search (BFS) to explore and collect connected nodes until the desired number of samples (e.g., 1000) is reached."

    The paper's stated goal is to predict links for isolated nodes, but the training ground truth is defined in Section 3.3 as '1-degree nodes and their neighbors,' and the test subgraph is sampled exclusively from the largest connected component. Thus the 'isolated-node' evaluation contains no isolated nodes at all; every evaluated node has degree at least 1. The claimed result—'the modified GraphSAGE outperforms the baseline model in predicting co-purchasing links for isolated nodes'—is therefore a renaming of connected-node edge prediction. No truly unseen or isolated product ever appears in the evaluation, so the new-product recommendation claim is not derived from the experiment; it is the same connected-component link prediction relabeled as isolated-node prediction.

  2. self definitional [Section 3.5.2]
    "For pairs of nodes, we sample from the target node's neighbors and use these as a proxy for aggregation. ... By introducing these modifications, our approach ensures that GraphSAGE can still be applied to isolated nodes in the co-purchasing network, leveraging the local neighborhood information and graph structure to make relevant predictions even in the absence of direct neighbors."

    An isolated node is defined earlier in the paper as a node with no direct neighbors. The proposed method for handling isolated nodes is to sample from the target node's neighbors, but for an isolated node that set is empty by definition, so the sampling procedure cannot be executed. The claim that the modification 'ensures that GraphSAGE can still be applied to isolated nodes' is therefore self-definitional: the method's required input (a non-empty neighbor set) is exactly what the target population lacks. Any reported performance on isolated nodes cannot come from the described procedure, because the procedure is undefined for degree-0 nodes.

full rationale

The paper contains no load-bearing self-citations and no imported uniqueness theorem; the references to GraphSAGE, Sentence-BERT, and Random Forests are standard external methods. However, the central claim—an inductive recommendation system for newly listed, isolated products—is not tested on isolated products. Section 2.2 removes all 159,575 isolated nodes by taking the largest connected component, and Section 4.3 evaluates on a BFS-connected subgraph of that component, so every test node has at least one neighbor. The training positives in Section 3.3 are existing edges incident to 1-degree nodes in that connected component, so the model learns to reproduce observed co-purchase edges among connected nodes. The reported top-k accuracy (GraphSAGE 0.0187 vs Random Forest 0.0125 at top-5) is therefore evidence only for edge prediction on connected nodes. Calling this a solution for isolated new products is a relabeling of the measured task, not a derivation of isolated-node performance. Additionally, the Section 3.5.2 remedy for isolated nodes—sampling from the target node's neighbors—is undefined for the very nodes it claims to serve. These two issues make the central 'new product' result partially circular: the evaluation target is constructed to coincide with the training distribution and excludes the population named in the claim.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on several domain assumptions about what constitutes a good proxy for new products and how co-purchase semantics translate to graph structure. The free parameters are mostly arbitrary or undisclosed, and no invented entities are introduced.

free parameters (6)
  • PCA dimension for title embeddings = 32
    Chosen to reduce title embedding dimensionality; no justification or sensitivity analysis.
  • Number of positive training samples = 10000
    Arbitrary choice; no analysis of how this affects performance.
  • Number of negative training samples = 10000
    Arbitrary choice; no analysis of class balance effects.
  • Number of trees in random forest baseline = 100
    Default setting; no tuning reported.
  • BFS subgraph sample size for evaluation = 1000
    Chosen as 'desired number of samples'; no justification.
  • GraphSAGE hyperparameters (learning rate, hidden size, epochs)
    Not reported in the paper; these are free parameters implicitly chosen by the authors but not disclosed, making the results non-reproducible.
assumptions (5)
  • domain assumption 1-degree nodes are representative proxies for isolated nodes in link prediction
    Section 3.3 uses 1-degree nodes and their neighbors as training data for the isolated-node problem, but no validation is provided that degree-1 structure resembles degree-0 structure.
  • domain assumption Co-purchase relationships are undirected
    Section 2.2 states that co-purchase is mutual, so they treat the graph as undirected. This is a reasonable domain assumption but still an assumption.
  • domain assumption One-hop neighbors are sufficient; two-hop co-purchases do not imply links
    Section 3.5.2 restricts sampling to one-hop neighbors based on the claim that indirect co-purchase does not imply a direct relationship. This is a modeling choice, not empirically validated.
  • domain assumption Category similarity function captures meaningful relatedness
    Algorithm 1 constructs a binary vector marking the deepest common category, but the paper does not validate that this similarity aligns with actual co-purchase behavior.
  • domain assumption Removing products with missing information does not bias the graph
    Section 2.1 says products with missing information were removed, but the extent and potential bias of this filtering are not discussed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Building a Recommendation System Using Amazon Product Co-Purchasing Network." pith.science (2026). https://pith.science/paper/U6SGN4IE

@misc{pith2026250602482,
  author       = {Pith},
  title        = {Pith review of: Building a Recommendation System Using Amazon Product Co-Purchasing Network},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U6SGN4IE}},
  note         = {Machine review of arXiv:2506.02482}
}
read the original abstract

This project develops an online, inductive recommendation system for newly listed products on e-commerce platforms, focusing on suggesting relevant new items to customers as they purchase other products. Using the Amazon Product Co-Purchasing Network Metadata dataset, we construct a co-purchasing graph where nodes represent products and edges capture co-purchasing relationships. To address the challenge of recommending new products with limited information, we apply a modified GraphSAGE method for link prediction. This inductive approach leverages both product features and the existing co-purchasing graph structure to predict potential co-purchasing relationships, enabling the model to generalize to unseen products. As an online method, it updates in real time, making it scalable and adaptive to evolving product catalogs. Experimental results demonstrate that our approach outperforms baseline algorithms in predicting relevant product links, offering a promising solution for enhancing the relevance of new product recommendations in e-commerce environments. All code is available at https://github.com/cse416a-fl24/final-project-l-minghao_z-catherine_z-nathan.git.

Figures

Figures reproduced from arXiv: 2506.02482 by the authors.

Figure 1
Figure 1. Plot of log connected component size with log frequency [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Degree Distirbution Scatter Plots top 50 highest-degree nodes, all located within the LCC, along with their neighbors, and imported the subgraph into Gephi. By zooming in on a specific section of the graph as showed in 3, we observed a local bridge connecting two distinct communities [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Zoomed in visualization of top 50 nodes and their neighbors. Node color represents the group [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Top-k accuracy for Modified-GraphSAGE, Random Forest, and Random Choice. Top-5 Accu [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: A sampled visualization for predictions in a subgraph with 100 nodes: Dark blue edges are [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

4 extracted references · 3 canonical work pages

  1. [1]

    The dynamics of viral marketing,

    J. Leskovec, L. A. Adamic, and B. A. Huberman, “The dynamics of viral marketing,” ACM Trans. Web, vol. 1, p. 5–es, May 2007

    work page 2007

  2. [2]

    Inductive representation learning on large graphs,

    W. L. Hamilton, R. Ying, and J. Leskovec, “Inductive representation learning on large graphs,” 2018

    2018

  3. [3]

    Sentence-bert: Sentence embeddings using siamese bert-networks,

    N. Reimers and I. Gurevych, “Sentence-bert: Sentence embeddings using siamese bert-networks,” in Proceed- ings of the 2019 Conference on Empirical Methods in Natural Language Processing, Association for Compu- tational Linguistics, 11 2019

    work page 2019

  4. [4]

    Random forests,

    L. Breiman, “Random forests,” Machine Learning, vol. 45, no. 1, pp. 5–32, 2001. A Extract Category Similarity Function Let C1 and C2 be the category lists for nodes node 1 and node 2, respectively. The goal is to compute the category similarity for each pair of nodes (node 1, node2). Algorithm 1 Extract Category Similarity 1: Input: samples, category inde...

    work page 2001

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.