Recommender System Study Notes - Nomination(4/6)

• Two tower model full update/incremental update.
  • Online Nomination: Online compute user ID, ANN in stored vector database with item.
  • Full Update: 1 epoch, 1 day.
  • Incremental Update: Only upadte ID embedding, frozen the fully connected layer.
• Two tower model Self-supervised learning
  • Problem: Two tower cannot learn item with limited amount of clicks (data biased).
  • Self-supervised learning: random feature transformation
  • bi’, bi’’ high similarity
  • bi’, bj’’ low similarity
  • Training:
    • pick n (user - item) as a batch for training.
    • m items as a batch(uniform sampling) for self-supervised learning.
    • $\frac{1}{n} \sum_{i=1}^{n} L_{\text{main}}[i] + \alpha \cdot \frac{1}{m} \sum_{j=1}^{m} L_{\text{self}}[j]$

Online Nomination

Implementation

• Item
  • Use neural network to compute item b’s embedding.
  • Store them in vectorized databse (Milvus, Faiss, HnswLib).
  • Indexing, for ANN search.
• User
  • Use user ID, feature, etc to compute user a’s embedding.
  • ANN search
    • Use a as query.
    • Return the top k as return.
  • Why store item but compute user online?
    • User is only 1, item is O(b).
    • User often changes quickly.

Model Update

Full Update VS Incremental Update

• Full Update
  • Use yesterday’s tfrecord, train 1 epoch.
  • Update the new user neural network, store item data into database.
  • Easy to implement, easy requirement for the system.
• Incremental Update(online learning)
  • Why? User’s interest can change quickly.
  • Data streaming, TFRecord.
  • Online learning, gradient descent, only update ID Embedding, not other parameters in the neural network.
  • Update the new ID embedding.
  • To be discarded after full update.
• Why only incremental update is not having the best result?
  • Hourly data is biased.
  • Full update: random shuffle, 1 epoch.
  • Incremental update: train from morning to night, this is not having the best result.
  • random shuffle makes a positive impact on the learning.

Two tower model Self-supervised learning

Problem

• The head effect in the recommendation system is severe:

  • A small number of items account for most of the clicks.
  • Most items have a low number of clicks.

• The features of high-click items are well-learned, but the features of long-tail items are not well-learned.

• Self-supervised learning: data augumentation, to better learn long-tail items’ embedding.

Intuition

• A simple feature transformation will make minimal impact on their feature embeddings, and pairwise, the similarity still remains large.
• i -> i’ -> bi’, i -> i’’ -> bi’’, j -> j’ -> bj’, j -> j’’->bj''
• i, bi’ and bi’’ has high similarity
• i, j, bi’ and bj’’ has low similarity
• cos(bi, bi’’) big, cos(bi’, bj’) small

Methods

1. Random Mask

• Randomly pick labels, and set them to null.

2. Dropout

• Randomly dropout 50% label.

3. Complementary

• ID, Label, Keywords, City.
• Randomly divide them into two {ID, keyword} and {Label, City}.
• {ID, default, Keyword, default} -> Item embedding.
• {default, Label, default, City} -> Item embedding.

4. Mask a set of correlated features

• U = {Male, Female, Unisex}.
• V = {Photography, Football, IT, Makeup}.
• u = female, v = Makeup, p(u, v) is high.
• u = female, v = IT, p(u, v) is low.
• $MI(U, V) = \sum_{u \in U} \sum_{v \in V} p(u, v) \cdot \log \frac{p(u, v)}{p(u) \cdot p(v)}$

Learning:

• Assume there are a total of \( k \) features. Calculate the MI (Mutual Information) between any two features offline, resulting in a \( k \times k \) matrix.
• Randomly select one feature as a seed and find the \( k/2 \) features most related to the seed.
• Mask the seed and its related \( k/2 \) features, retaining the remaining \( k/2 \) features.

Training:

• Sampling m items as a batch
• cos(bi’, bi’’) -> 1, other cos(bi’, bj’’) -> 0