Home page for the SDSR project
Revise your presentation on paper P using the feedback from reviewers. Prepare a 500 word narrative summarizing the revisions for suggested improvements from reviewers. Place the revised presentation and narrative on the web and ensure that the course web-site has a link to your material.
The feedback for our slides was fairly positive overall. There were, however, a few improvements that the feedback suggested. Noah revised the slides based on their feedback, and then we reviewed and finalized the design of the edited slides together.
Both reviews asked for more examples and/or imagery to help relate the ideas to real-world scenarios. To address this, Noah added wildfire prediction as an example of an objective that can be approached with either homogeneous data or hetereogeneous data. The wildfire example included geographic maps for drought data and vegetation data, and he included text that indicated using both datasets would be heterogeneous while only using one would be homogeneous. Noah also added an image of a chart to the Contributions: Heterogeneity and an image of various homogeneous and heterogeneous graph examples to a new Homogeneous vs Heterogeneous slide. I suggested we move the graph examples to the heterogeneity slide since the context was still fitting and the presentation flow was improved.
One critique brought up from Group 3 was to relate possible applications to real-world applications. In response, Noah added a slide about Relational Deep Learning (RDL) and how RDL connects to graph learning. In addition, he provided the MovieLens application as an example of a recommender system built by RDL. The example included a photo of members of the team, and a glance of the website hosting the application.
Both reviews asked for more specific details across the following topics: message passing, the explainer callback mechanics, and temporal graph sampling. In response, Noah added short bullet-points about the EdgeIndex tensor, PyG’s temporal graph sampling, and the Explainer callback mechanism within the message passing process.
Submit two questions to test the understanding of key concepts in P in a closed-book examination. Attach model solutions to each questions. Each question should be answerable in 10 minutes. Questions should be problem-solving in nature with a unique answer. Questions can be similar to the quiz from the presentation. If similar, provide a 75-100 word justification of the answer, explaining why the choices are correct or incorrect.
You are building a GNN to predict large wildfires in the United States. You have the following data sources:
You must design a heterogeneous graph schema for PyG 2.0 that best captures causal structure for prediction. Which of the following node/edge schema is most appropriate for a scalable heterogeneous GNN in PyG 2.0?
(A) A single node type, “Region”, with all features as node attributes and no edges.
(B) Node types of “Region” and “Wildfire” with edges of type Region–Region for adjacent regions, and Region–Wildfire when wildfires occurred in region. The measures of drought, vegetation and climate type will be regional node features.
(C) Node types of “Drought”, “Vegetation”, “Climate” and “Wildfire”. Edges between nodes such as Wildfire-Climate and Wildfire-Wildfire represent shared timestamps when present.
(D) A single node type of “Wildfire” with edges between wildfires in chronological order. All other data will be node features.
Answer: (B).
Reasoning: Option B cleanly separates entities (regions, wildfires) into node types and uses edges to encode meaningful relations: spatial adjacency (Region–Region) and event occurrence (Region–Wildfire). This matches PyG 2.0’s heterogeneous and temporal capabilities, where “entities become nodes, primary-foreign key links become edges” and “heterogeneous graph data types, message passing… temporal graph handling” are first-class. By placing drought, vegetation, and climate as features on “Region” nodes, the model can learn causal patterns across space and time while remaining scalable and structurally interpretable.
Meanwhile, Option (A) has no edges and is thus not a graph system. Option (D) is a homogeneous graph structure. Option (C) fails to connect any nodes without a time component, such as climate and vegetation.
A PyG 2.0 explainer is applied to a social-network GNN that predicts whether a user is likely to commit fraud. The explainer outputs the following details:
The edge between **Tom** and **Robert** is highlighted as highly relevant
For **Tom**, the feature interest = "bitcoin" is marked highly relevant; location = Istanbul is marked irrelevant
For Robert, the feature interest = "crime", "fraud" is marked highly relevant
You are told that the explainer works by learning a differentiable mask over edges and node features, and then re-running the model with masked messages to see if the prediction changes as little as possible.
1) Which substructure attribution technique is being used?
2) Which aspects of the GNN are directly exploited by this explainer? Choose the most complete option.
Answers: (C) and (C).
Reasoning:
(1) Key descriptions such as “learning a mask,” “masking messages,” and “re-computing the forward pass without changing the output” are part of GraphMask. GraphMask uses vertex hidden states and messages at layer ‘k’ to predict a mask and re-compute the forward pass with modified node states. Feature Importance (A) and SHAP (B) typically operate on features or inputs, not learned message masks. Counterfactual reasoning (D) would explicitly change outcomes.
(2) Options (A), (B) and (D) are disqualified since the mask is learned over both nodes and edges. The explainer leverages hidden states and messages in the message-passing layers, so (C) is uniquely correct.
Midterm Slides (example):
The slides should summarized Heilmeir Questions such that each questions should be addressed in 1-2 slides. The questions are as follows:
- Problem Statement (1 slide): What is the group trying to do?
- Significance of the problem (1 slide) : If the group succeeds, what difference does the group think it will make?
- Challenges (1 slide): Why is the problem hard?
- Related Work and its limitations (1-2 slides): How does the problem get address, at present? What are the limitations of the state-of-the-art approaches?
- Proposed Approach (1-2 slides): What is the new technical idea? How does it overcome the limitations of the state-of-the-art approaches? In what sense, is the proposed approach novel? In what sense, is it better than the state-of-the-art?
- Choice of validation methodology (experimental, analytical) (1-2 slides): Which validation methodology will the group use to evaluate the proposed approach? Choices include example, experiments, analytical methods, case study, detailed illustrative examples, prototyping and demonstration of new capability, etc. List key steps within chosen methodologies. You may look at published papers using the chosen methodology to identify these steps. For example, an experimental methodology may include a description of the experiment design listing the candidates to be compared, metrics of performance, values of fixed parameters, value sets for variable parameters, benchmark datasets and computations, key assumptions, etc.
- Conclusions and Future Work (1 slide).
- Weekly Plan and Task (1 slide): This slide should contain a table with 5 rows and 2 columns, where the column names are ‘Weeks’ and ‘Task’. Each row will be filled with a ‘Week’ (e.g., 3/3 - 3/10) and a ‘Task’ (e.g., running experiments, etc.).
Formal Proposal:
The proposal is a brief document describing the components listed in CP4 using narrative, pictures, tables, etc. You may follow the format of this example from previous year.
Peer-review two other drafts, using the Digital Accessibility Resource Guide and Accessibility Checker. Write a peer-review summary and submit it. (Team G1 will peer-review teams G2 and G3 / G2 will peer-review G3 and G4 / … / G8 will peer-review G9 and G1 / G9 will peer-review G1 and G2).
G3 Review
G3, as of February 21, 2026, has not presented a summary of the edits they made to their assigned slides. We were able to download a copy of their slides from their webpage on that date and review them using the Microsoft ‘Review’ feature and checking it against the Digital Accessibility Resource Guide.
Initializing the ‘Review’ feature suggests that they likely did not use the feature during their DA2. The feature found: 33 instances of hard-to-read text, 189 missing alt text, 2 missing table headers, 20 missing slide titles, 23 duplicate slide titles, and 48 reading-order checks to verify. They did not mark 97 objects in the slides as ‘decorative’, though based on reviewing those objects they were all intended to either hide or animate a key feature on the slides.
Based on the ‘Review’ findings, G3 likely did not finish their slides’ digital accessibility updates before we reviewed them.
G4 Review
G4 provided before-and-after images to show the progress of their updates in DA2, as well as a downloadable version of their slides on their webpage. Reviewing our copy, downloaded on February 21, 2026, the ‘Review’ feature confirmed that they had marked off most of the updates it suggested. The only remaining suggestions were reading-order checks, but we could not find any issue the 40 orders suggested.
The title numbering seemed off for the “Outline for LO1” and “QPO Challenges in SDBMS” slides. They chose to put the numbers at the front of the titles, which suggests that there are sections with different subjects that are connected rather than a single subject displayed across multiple slides. We believe it would be clearer to connect the slides by using “- 1” and “- 2” in the titles, as they did for the “Community Variability in Review” slides and “Structure of a referee report” slides. Updating the “L01” and “QPO” slides in this manner would also help keep the presentation of the slides consistent. We also found the “How not to evaluate a paper - Do not do” title to be a bit off, since the part before the “-“ seemed fine on its own.
There were also a lot of titles that could also connect to their subject matter more clearly. There were several “Quiz”, “Limitations of Related Work”, and “Learning Objective” slides that could have slight title changes to connect the slides to their respective, specific content. Slide 12 also appears to have the same title as the slide after it: “Taught to the Test”. The first slide appears to have an image with the name in the title positioned high enough to be treated as a title. Rearranging the image and giving a slightly different titles for each slide to show they’re connected would help alleviate some of the confusion over the slide titles. They could also leave the first slide as is and rename the second one to, “Taught to Test - 2”, to keep in line with the other ‘continued’ slides’ shared format.