基于时空与事件类别数据集的时空事件预测模型选型问询

Hey there! Let’s walk through your three prediction goals and pick the best-fit models for your event dataset—since you noted existing solutions don’t align with your scenario, I’ll focus on approaches that play to your data’s strengths (lat/lon, timestamp, event categories, and only actual occurrences):

1. Predict Next Event Location Given Current Timestamp #

This task centers on learning temporal sequences of spatial events, so models that fuse time and space dependencies work best:

• Spatio-Temporal LSTM (ST-LSTM)：Built explicitly for spatio-temporal sequence data, this model captures both the sequential patterns of events over time and spatial correlations between locations. You can feed event categories as additional input features to refine predictions (e.g., certain categories might have more predictable location shifts).
• ST-Transformer：If your dataset has long-term temporal patterns (e.g., events recurring every few months in specific areas), the transformer’s attention mechanism excels at picking up these long-range spatio-temporal dependencies. It’s especially useful when you have a large volume of historical event data.
• Hidden Markov Model (HMM)：If your events follow clear state-based location transitions (e.g., category A events often move from Location X to Y), HMM treats geographic regions as "states" and learns transition probabilities over time. It’s lightweight, easy to implement, and has strong interpretability.

2. Predict Most Likely Event Time Given Current Location #

Here, we’re modeling spatial-to-temporal patterns—focusing on how event timing correlates with specific locations:

• Self-Exciting Point Processes (SEPP)：A type of Temporal Point Process (TPP) designed specifically for event timing. It models the likelihood of an event occurring at a location based on past events there (e.g., a surge in events at a spot might increase the chance of another event soon). You can tailor parameters per event category to make predictions more precise.
• XGBoost/LightGBM (with Temporal Feature Engineering)：Treat this as a regression (predict time until next event) or classification (predict time window) task. Extract features like: average time between past events at the location, hour/day/week/month of past occurrences, and holiday flags. Tree-based models handle these structured features well, are fast to train, and easy to interpret.
• LSTM for Time Intervals：Feed the sequence of past event timestamps at the target location into an LSTM. It learns periodic patterns (e.g., events at Location Z always happen every 7 days) and irregular time gaps, making it great for capturing non-linear temporal dependencies.

3. Predict Event Occurrence Probability Given Time & Space #

This is a spatio-temporal density prediction task, where we need to model probability across time and location:

• Spatio-Temporal Gaussian Processes (ST-GP)：This model explicitly models both spatial correlations (e.g., adjacent locations have similar event probabilities) and temporal correlations (e.g., a location’s event probability rises every weekend). It outputs a full probability distribution, which is exactly what you need, and has strong interpretability for small-to-medium datasets.
• U-Net-based Spatio-Temporal Segmentation：If you discretize your data into spatial grids (e.g., 100m x 100m map tiles) and time bins (e.g., hourly slots), a U-Net variant can learn to predict event probabilities for each grid-time bin. It’s ideal for large-scale datasets, as it captures both local (neighboring grids) and global (entire map) spatio-temporal patterns. You can also incorporate event categories as additional input channels.
• Spatio-Temporal Random Forest：Extend traditional random forests by adding spatio-temporal features (e.g., number of events in neighboring grids in the past 24 hours, event count in the same grid over the past week). It’s computationally efficient, handles mixed data types well, and gives you feature importance scores to understand what drives event probabilities.

Quick General Tips #

• Feature Engineering is Key: Split timestamps into hour, day of week, month, and holiday flags; convert lat/lon into grid IDs or distance to high-event locations; encode event categories as embeddings or one-hot vectors.
• Handle Imbalanced Data: Since you only have actual events (positive samples), generate negative samples by randomly selecting spatio-temporal points where no event occurred. This helps models learn to distinguish between event and non-event regions/times.

内容的提问来源于stack exchange，提问作者Mido