Contrastive Learning Perspectives on Retrieval and Ranking Models in Recommendation Systems

The speaker introduces contrastive learning from a historical and technical viewpoint, tracing its roots to metric learning and self‑supervised methods like BERT, and describing its core pipeline: automatic positive‑sample generation, random negative sampling, encoder mapping, and the InfoNCE loss that enforces alignment of positives and uniformity of embeddings.

Key components of a contrastive learning system are identified as (1) how positives are constructed, (2) how the encoder maps inputs to a projection space, and (3) how the loss function is designed. Variations in these components lead to different models.

Typical image‑domain contrastive models are presented:

SimCLR uses in‑batch negatives, data augmentations for positives, a ResNet encoder plus a projector, L2‑norm on embeddings, and InfoNCE loss with a temperature hyper‑parameter.

MoCo introduces a momentum‑updated encoder and a large negative queue to overcome batch‑size limits.

SwAV combines contrastive learning with clustering, assigning each view to a prototype and optimizing similarity to the prototype.

The speaker then maps these ideas to recommendation systems, arguing that dual‑tower retrieval models are essentially contrastive learning systems. The dual‑tower architecture splits user and item features, encodes them separately, and computes similarity (inner product or cosine). Practical decisions include:

Negative sampling: in‑batch negatives, global random negatives, or a mix of both to mitigate selection bias.

Embedding normalization: applying L2‑norm (or using cosine similarity) improves training stability and linear separability.

Temperature scaling: a small temperature focuses the loss on hard negatives, similar to focal loss, and yields significant performance gains.

These three practices are explained through the lens of alignment and uniformity: in‑batch negatives increase uniformity, temperature emphasizes hard negatives, and normalization enforces consistent embedding magnitudes.

Beyond standard dual‑tower models, the speaker proposes extensions:

Adding contrastive auxiliary losses on the item side (e.g., dropout‑generated views) to strengthen long‑tail item embeddings.

Applying the same auxiliary loss on the user side, possibly using user behavior sequences or graph neural networks.

Integrating graph neural networks (GNNs) for graph‑based retrieval, where sub‑graph augmentations (node dropping, edge perturbation, feature masking, random walks) create contrastive views.

Finally, the talk suggests that contrastive learning can also be incorporated into ranking models, provided suitable positive‑sample constructions and loss functions are defined.