Failure Modes: False Negatives and Label Noise

False negatives are the most damaging failure mode in hard negative mining. At web scale, near duplicates or items with the same semantic label often enter the training data as negatives due to noisy labeling or incomplete metadata. When hard mining selects these false negatives because they are close to the anchor, the loss aggressively pushes them apart. This harms recall for near duplicates and creates adversarial separation of items that should cluster together. For example, in product search, two listings for the same product from different sellers might be mislabeled as negatives. Hard mining would teach the model to separate them, degrading deduplication quality. Mitigation starts with metadata filtering. Before mining, remove candidates that share critical identifiers like product SKU, user ID in social networks, or content hash in media systems. In face recognition, filter negatives from the same identity using known identity labels. For text or multimodal data, use approximate string matching or perceptual hashing to detect near duplicates and exclude them from negative pools. Supervised contrastive loss that supports multiple positives per anchor is more robust because it averages over positives, reducing the impact of any single mislabeled positive or negative. Label noise amplifies when mining focuses on hard examples. If 5% of labels are incorrect, hard mining disproportionately selects these mislabeled pairs because they violate expected distance patterns. This can cause exploding loss or model collapse as the optimizer chases contradictory signals. The solution is a curriculum schedule that starts with semi hard or even random negatives for the first 20 to 30% of training, then gradually increases hardness. Gradient clipping at norm 1.0 to 5.0 prevents single bad batches from destabilizing parameters. Some systems use noise robust objectives like bootstrapping, where the model's own predictions smooth noisy labels over time. Batch bias is another edge case. In batch mining only explores negatives within the current batch, so poor shuffling or small batch sizes limit hardness. With a batch of 128 and 10 classes, each anchor sees at most 12 negatives per class on average. If the dataset has 1000 classes, most classes never appear as negatives for a given anchor. Monitor active triplet fraction per batch: if it drops below 10%, batches are too easy. Use cross device gathering to enlarge the negative pool, or increase batch size. Stratified sampling that ensures each batch has diverse classes helps, but adds data pipeline complexity.

E commerce product search: Two sellers list the same product with different titles. Without SKU based filtering, hard mining treats them as negatives. The model learns to separate identical products, breaking deduplication. Solution: Filter negatives where SKU matches or image perceptual hash distance is under threshold 5.

Face recognition with label noise: 2% of face images are mislabeled to wrong identity due to annotation errors. Hard mining selects these mislabeled faces because they cluster with wrong identity. Curriculum schedule uses semi hard mining for first 10 epochs, then transitions to hard mining. Gradient clipping at norm 2.0 prevents loss spikes.

Music recommendation: User listening history contains accidental clicks (user skipped within 5 seconds). These create false positive engagement signals. Hard negative miner retrieves songs user skipped, but some are false negatives (user actually liked but was interrupted). Filter negatives where play duration exceeded 30 seconds to reduce false negative rate from 8% to 2%.