Smart Ways to Improve Learning in Time-Based Networks

Time-based networks, such as social media or online shopping platforms, illustrate how connections evolve over time. Recently, scientists have been focused on developing better models to understand these networks. However, there's a significant oversight: the quality of the "wrong" examples used to train these models has often been neglected.

Key Challenges in Time-Based Networks

1. Imbalance of Examples:

   • There are far fewer "right" examples compared to "wrong" ones at any given time.
   • The "right" examples also change over time.

2. Solution: Curriculum Negative Mining (CurNM):

   • A new approach designed to adjust the difficulty of "wrong" examples as the model learns.

How CurNM Works

• Diverse Pool of "Wrong" Examples:
• Mixes different types of examples to compensate for the scarcity of "right" ones.
• Focus on Recent Changes:
• Captures shifting patterns in the network.
• Randomness for Stability:
• Ensures stable training by introducing controlled randomness.

Results and Validation

• Testing:
• Evaluated on 12 datasets and three different models.
• Outcome:
• Proven to be more effective than existing methods.
• Further experiments confirmed its reliability and usefulness.