Speak clearly about your research
A practical script for an English interview on spatio-temporal data mining, urban computing, and your LLM-agent mobility project.
Use the blue boxes as spoken answers. The side notes are coaching prompts, not text to memorize word for word. Switch to EN only for the final rehearsal.
Introduce yourself without overclaiming
Fill in the bracketed facts once. Keep the structure: background -> research interest -> relevant project -> why this RA.
“Hello, I’m [your name]. My background is in [degree / field], and I’m especially interested in how machine learning can model structured data that evolves across space and time. Recently, I worked on a project about next-location prediction with an LLM-driven agent. The project made me interested in evidence-grounded reasoning, urban mobility, and reliable evaluation. I’m applying for this RA role because I want to deepen my understanding of spatio-temporal data mining and contribute to research that is both technically sound and useful for cities.”
“Urban data is interesting to me because it is structured but not simple. A location is related to nearby locations, and the same location behaves differently at different times. That means a useful model has to capture spatial dependence, temporal dependence, and uncertainty at the same time. I also like this area because the results can connect directly to transportation, environmental monitoring, and urban planning.”
“I see a strong fit between my interests and the group’s work on spatio-temporal data mining, urban computing, and smart-city applications. The papers I read also showed me that the field is moving from task-specific models toward transferable representations and more reliable decision systems. I would like to learn from that research environment and contribute through careful data processing, modeling, experiments, and communication.”
Explain AgentMob in 90 seconds
One-line summary
Memorize: “AgentMob formulates next-location prediction as an adaptive, evidence-controlled decision process. It uses a fast path for routine cases and selective tool calling for ambiguous cases.”
Structured project pitch
Four tools: say the purpose, not the implementation
| Tool | Plain-English purpose | Useful sentence |
|---|---|---|
| Mobility Context Retriever | Recent locations plus same weekday/hour history. | “It distinguishes immediate continuity from a long-term temporal pattern.” |
| Geographical Information Retriever | Distance and urban-function description for candidates. | “It checks geographical plausibility and semantic consistency.” |
| Stay-Move Estimator | Whether the person is likely to remain or depart. | “It is useful at the boundary between staying and moving.” |
| Historical Behavior Retriever | Dwell time, visit frequency, transitions, and regularity. | “It verifies whether a tempting candidate is supported by actual visits.” |
Metrics in one breath
“Acc@1 asks whether the first prediction is exactly correct. MRR@5 rewards putting the correct location near the top five, with reciprocal-rank weighting. Geographic distance measures how far the top prediction is from the ground truth, so a geographically close mistake is less severe than a distant one.”
Questions the professor may ask
For each question, answer in 20-60 seconds, then stop. A clear pause is better than adding unsupported detail.
Q1. What is the main contribution?
Adaptive evidence control: fast path for routine cases, iterative tools for ambiguous cases, with auditable traces.
Q2. Why not use a supervised Transformer?
It is a strong choice when labeled, stable, task-specific data are available. AgentMob targets training-free adaptation and decision transparency, not universal superiority.
Q3. Why is the fast path important?
Many mobility instances are repetitive. Calling an LLM for every sample wastes tokens and latency; difficulty-aware computation is more efficient.
Q4. What if evidence conflicts?
Keep multiple candidates, query targeted tools, compare temporal, transition, stay-move, and spatial evidence, and stop only when confidence is sufficient.
Q5. How did you prevent leakage?
Use chronological train/test splits; for every target, tools access only training records and observations before the target timestamp.
Q6. What is a weakness?
Tool boundaries and invocation policy are manually specified; smaller LLMs may fail to call tools reliably; semantic information is weak in anonymized grids.
Q7. Does explainability mean correctness?
No. A fluent rationale can still be wrong. Audit traces expose evidence, but calibration and faithfulness still need evaluation.
Q8. How would you generalize to a new city?
Separate city-specific preprocessing from the controller; test zero-shot and few-shot transfer; normalize spatial units; add local POI/road knowledge; report performance by city and regime.
Q9. What would you improve?
Learn the tool-selection policy, add calibrated uncertainty, evaluate privacy-preserving deployment, and compare with STGNN or CityFM representations.
Q10. How do the readings connect?
STGNN models graph structure; SSL learns from unlabeled sequences; CityFM learns reusable geospatial embeddings; AgentMob uses grounded evidence and an LLM controller.
Follow-up answers that sound mature
Short answers for fundamentals
| Prompt | Answer to memorize |
|---|---|
| What is spatio-temporal data? | “It records variables associated with locations and time. The observations are usually correlated across nearby locations and across time, and the correlation can change by region or period.” |
| What is a graph neural network? | “A GNN updates each node by aggregating information from its neighbors according to the graph structure. This lets the model use relations that a plain sequence model would ignore.” |
| GCN vs GAT? | “A GCN usually applies a normalized, structure-based aggregation. A GAT learns attention weights so different neighbors can contribute differently.” |
| Why use a TCN? | “A causal TCN uses convolutions over past values only. Dilations enlarge the receptive field and allow parallel computation over a history window.” |
| What is self-supervised learning? | “It creates a proxy label from the data itself, such as predicting a masked segment or matching two augmented views, to learn a representation before the downstream task.” |
| What is contrastive learning? | “It pulls representations of positive pairs together and pushes negative pairs apart. The difficult part is constructing pairs that reflect the right invariances.” |
| What is a foundation model? | “It is a broadly pretrained model or representation that can be adapted to many downstream tasks. The key idea is reuse and transfer, not just model size.” |
| Why does geography need grounding? | “Coordinates and urban functions are not automatically reliable in an LLM. We need spatial encodings, retrieved facts, structured descriptions, or tools to connect language reasoning to real geography.” |
| What is distribution shift? | “The data distribution changes between training and deployment, for example across cities, seasons, policies, or sensor systems. A model can lose accuracy even when the task definition stays the same.” |
Connect your own project to the readings
必须记忆的表达
Structure your answer
At a high level, ...
从高层次来说……
The key design choice was ...
关键设计选择是……
The intuition is that ...
直觉是……
Concretely, the pipeline is ...
具体流程是……
The result suggests ..., rather than proving ...
结果说明……，但不能证明……
Discuss evidence and limits
To avoid temporal leakage, ...
为了避免时间泄漏……
An important caveat is ...
一个重要限制是……
This is dataset-dependent.
这取决于数据集。
I would validate this with an ablation / stress test.
我会用消融 / 压力测试验证。
I would not generalize beyond the reported setting.
我不会把结论外推到论文没有测试的情形。
| Say this | Instead of this | Reason |
|---|---|---|
| “I worked on a project.” | “I did a project.” | More natural professional English. |
| “The model predicts the next location.” | “The model prediction the next location.” | Third-person singular needs “predicts.” |
| “We evaluated the model on three datasets.” | “We tested in three datasets.” | Use “on” for datasets. |
| “The method is trained with unlabeled data.” | “The method is training no label data.” | Use a complete passive construction. |
| “I was responsible for preprocessing and evaluation.” | “I responsible for ...” | Past tense requires “was.” |
| “This result is promising, but it has limitations.” | “This result is very good.” | Balanced research language sounds more credible. |
Vocabulary to rehearse
spatial dependencytemporal dependencyspatial granularitytrajectorystay-move behaviorhistorical regularitycandidate locationgeographical plausibilityevidence-groundedadaptive tool callingchronological splittemporal leakagedistribution shiftcalibrated uncertaintyaudit traildownstream tasktask-specific baselineablation studygeneralizationlatency / token cost
Useful connectors
“There are two reasons for this.”
“The first is computational; the second is methodological.”
“In contrast, ...”
“This matters because ...”
“My current understanding is ...”
A compact practice plan
Personal facts to fill before the call
| Fact | Your exact wording |
|---|---|
| Degree / current status | [ ] |
| Role in the mobility-agent project | [ ] |
| One concrete technical contribution | [ ] |
| One experimental or debugging challenge | [ ] |
| One result you can defend | [ ] |
| Why this RA / why this professor | [ ] |
| One question to ask the professor | [ ] |
“What kinds of spatio-temporal data does the RA work with most often?”
“How do you balance methodological novelty with deployment constraints in the group’s projects?”
“For a new RA, what would a successful first three months look like?”
“Are there opportunities to work across graph models, foundation models, and urban applications?”