Abstract
Granularity and accuracy are two crucial factors for crime event prediction. Within fine-grained event classification, multiple criminal intents may alternately exhibit in preceding sequential events, and progress differently in next. Such intensive intent dynamics makes training models hard to capture unobserved intents, and thus leads to sub-optimal generalization performance, especially in the intertwining of numerous potential events. To capture comprehensive criminal intents, this paper proposes a fine-grained sequential crime prediction framework, CrimeAlarm, that equips with a novel mutual distillation strategy inspired by curriculum learning. During the early training phase, spot-shared criminal intents are captured through high-confidence sequence samples. In the later phase, spot-specific intents are gradually learned by increasing the contribution of low-confidence sequences. Meanwhile, the output probability distributions are reciprocally learned between prediction networks to model unobserved criminal intents. Extensive experiments show that CrimeAlarm outperforms state-of-the-art methods in terms of NDCG@5, with improvements of 4.51% for the NYC16 and 7.73% for the CHI18 in accuracy measures.
The Challenge of Fine-grained Crime Prediction
Predicting crime isn't just about identifying "hotspots"; it’s about understanding the "why" behind the "where." In the world of fine-grained crime prediction—where researchers try to guess the specific category of the next crime in a three-hour window—the biggest hurdle isn't a lack of data, but the sheer volatility of human intent. Most crime prediction models focus on volume (how much crime?) or general location. However, fine-grained prediction (what specific crime happens next at this exact spot?) is much harder due to what the authors call Intensive Intent Dynamics.
As shown in the figure above, at any given "spot" (like a specific bus stop or taxi stand), criminal intents (e.g., property vs. violence) switch frequently and progress differently. Standard deep learning models struggle here because they often fail to capture "unobserved intents"—the "dark knowledge" of what could have happened but didn't. Furthermore, trying to learn every possible crime category at once often leads to sub-optimal results because the model gets overwhelmed by the noise of rare or highly specific events.
The CrimeAlarm Framework
The core of CrimeAlarm is a model-agnostic distillation framework. It doesn't care if you use a Transformer, a GRU, or a TCN as your backbone; it changes how the model learns through three primary mechanisms:
- Online Mutual Distillation: Since there is no "perfect" pre-trained teacher model for crime, the authors use two "peer" networks that start from scratch and learn from each other's probability distributions.
- Decoupled Knowledge Distillation (DKD): The loss function is split into two parts: Target Distillation (learning from the crime that actually occurred) and Non-target Distillation (learning from the probabilities assigned to crimes that didn't occur, capturing the "dark knowledge").
- The Easy-to-Hard Pipeline (Curriculum Learning):
- Simple-to-Difficult Target Distillation: Early in training, the model focuses on "spot-shared" intents (common crimes). Later, it shifts focus to "spot-specific" intents that are harder to predict.
- Less-to-More Non-target Distillation: The model starts by only looking at a few frequent non-target crime categories. As training progresses, it gradually "unmasks" more categories, increasing complexity as the model becomes more capable.
Experimental Evaluation
The researchers evaluated CrimeAlarm on two massive, real-world datasets:
- NYC16: Crime records from New York City in 2016, featuring 3,229 spots and 440 distinct event classes.
- CHI18: Records from Chicago in 2018, featuring 2,692 spots and 246 event classes.
In both cases, time was divided into 3-hour slots, and "spots" were defined by a combination of precinct and premises (e.g., a grocery store in a specific precinct). CrimeAlarm consistently outperformed seven state-of-the-art baselines (including HAIL and CLC), improving NDCG@5 by 4.51% on the NYC16 dataset and 7.73% on the CHI18 dataset.
Key Insights and Future Work
A major takeaway is the value of "Dark Knowledge": the probability distribution over non-occurring crimes is just as important as the ground truth. If a model thinks a "robbery" was almost as likely as the "larceny" that actually happened, that's a vital clue about the underlying criminal intent of that area. Furthermore, because the distillation occurs during training, the model provides SOTA accuracy with the same inference speed and memory footprint as a standard single-network model during deployment.
While CrimeAlarm handles temporal dynamics well, the authors acknowledge that spatial modeling is currently implicit. Future work aims to integrate explicit spatial relationships (like graph neural networks) into the framework and expand the logic to other urban anomalies, such as traffic accidents or web behavior patterns.
Summary updated