Pattern Classification Model of Traffic Data including Missing Values

Research Motivation

---

Problems with Time Series Sensor Data

1. Missing values due to troubles in transmission and sensor defects
   → Needs proper processing before analysis

2. Requires time until meaningful amount of data is collected
   → Needs to accurately analyze small amount of data

3. Physical restrictions that put locational constraint of the sensors
   → Needs to analyze data that is affected by other variables

Suggested Model

---

Pattern Classification Model based on Attention Mechanism and CNN for Analysis of Traffic Data including Missing Values

1. Attention Mechanism
   Added an attention mechanism to focus on important entries in data
   Produces weights for each entries in data with an additional meta data input of whether each entry is missing

2. Deep model structure of CNN and Feed Forward layers
   Used deep model structure of CNN and Feed Forward layers to accurately classify minor pattern differences

Experimentation Data

---

• Used traffic data measured every 5 minutes on 101 North Freway, which is close to Dodger Stadium (consisting of 288 elements per day)
• Designed a model that classifies whether the stadium held a game that day from the day’s traffic data

Classification Result of Traffic Loop Time Series Data

---

• The result shows that the proposed attention mechanism improves the classification accuracy

Classification Result on Additional Traffic Data

---

• The result shows that the attention mechanism improves classification accuracy of data with missing values
• The result also shows that the attention mechanism can improve classification accuacy of data without missing values

Conclusion

---

 Conclustion

• Proposed an attention mechanism that can improve classification accuracy of time series sensor data with missing values without any complicated preprocessing procedure
• Through various experiments proved that the attention mechanism works to improve classification accurary with time series data with missing values

 Future Work

• Experiment with other types of time series data to see whether the mechanism can be applied to different sensor data
• Experiement with multiple layers of Feed Forward layer for attention mechanism