面向零售业务的Collaborative Filtering推荐系统：隐式评分偏差处理、替代指标及加权优化问询

Great question—this is a super common pain point when building implicit collaborative filtering systems for retail, where explicit 1-10 ratings just don’t exist. Let’s break down your two questions with practical, retail-focused solutions:

1. Better Implicit Rating Alternatives to Purchase Count/Views #

Purchase counts and page views are easy to track, but they’re far from the only signals that reflect user preference—especially for slow-moving, high-value products. Here are some stronger alternatives:

• Engagement Depth & Duration: For slow-moving items like TVs, a user spending 15+ minutes reading specs, checking reviews, and comparing models is a way stronger signal of interest than someone clicking a chip bag’s page for 2 seconds. Track metrics like time on product page, number of tabs/sections viewed (e.g., warranty details, installation guides), or whether the user shared the product link.
• Cart & Wishlist Signals: Adding an item to cart (even if not purchased immediately), saving it to a wishlist, or setting restock alerts are all indicators of intent that aren’t tied to purchase frequency. For big-ticket items, users often research, save, and come back later—these actions carry more weight than a single impulse buy of a snack.
• Repeat Category Interactions: Instead of counting purchases of a single slow-moving product, track how often a user interacts with the category (e.g., searching "4K smart TVs" multiple times, viewing 3 different TV models in a week). This captures sustained interest even if the user only buys one TV every few years.
• Contextual Purchase Signals: Look at the relative value of the purchase to the user’s overall spending. For example, if a user spends $800 on a TV (50% of their monthly retail budget) vs. $5 on chips (0.3% of their budget), the TV purchase should carry far more weight than 160 chip purchases.

2. Weighting Strategies to Fix Fast/Slow-Moving Product Bias #

If you do stick with purchase counts (or combine them with other signals), these weighting techniques will level the playing field:

• Inverse Popularity Weighting: Downweight signals from overpopulated fast-moving products and upweight slow-moving ones. A simple formula is:

  weight = 1 / log(1 + total_product_purchases)
  

  So a TV with 10 total purchases gets a weight of ~1/2.3, while a chip bag with 10,000 purchases gets ~1/9.2—making a single TV purchase worth ~4x a chip purchase in your implicit rating.
• User-Level Normalization: For each user, normalize their behavior relative to their own habits. For example, if User A buys 100 snacks and 1 TV, scale the TV’s count by the ratio of their average category spending:

  normalized_rating = (purchase_count / user_total_category_purchases) * category_weight
  

  Assign a higher category weight (e.g., 10 for electronics, 1 for snacks) to amplify slow-moving category signals.
• Time-Decay Adjustments: Slow-moving products reflect longer-term preferences, so adjust decay rates by category. For snacks, weight purchases from the last 30 days heavily (e.g., 1.0) and older ones lightly (0.2). For electronics, keep purchases from the last 6 months at full weight (1.0) since a user’s TV preference won’t shift overnight.
• Category-Level Quantile Scaling: Map each product’s purchase count to its percentile within its category. For example, if a TV is in the 90th percentile of TV purchases (i.e., one of the most bought TVs), give it a score of 0.9—same as a chip bag that’s in the 90th percentile of snack purchases. This ensures signals are comparable across categories, regardless of absolute purchase volume.

A pro tip: Combine multiple implicit signals into a single composite score first (e.g., 0.4*purchase_count + 0.3*cart_time + 0.2*page_duration + 0.1*wishlist_add) before applying these weighting strategies. This gives you a more holistic view of user intent.

内容的提问来源于stack exchange，提问作者Ad Astra