A Recommender System Algorithm Accounting for Cultural Factors to Solve Cold-Start Problems

Ekaterina Lemdyasova24 May 2025

Problem

A service platform has to recommend cloud or web services to a user who has just arrived and has no rating history. Standard collaborative filtering has nothing to work with, and approaches that fill the gap with profile or behavior data raise privacy concerns. The paper behind this talk asks how far one cheap, non-sensitive signal can go: the user's geographic location. The target is predicted quality of service (QoS): response time and throughput.

Example from the article: a new user from a given country gets predicted response time and throughput for services in the WS-Dream dataset (339 users, 5825 services), using only the country-level cultural profile of the user.

Solution in brief

Turn each country into a number, a Hofstede cultural distance, and use it in two places. First, find culturally similar users and services to fill the missing ratings of a cold-start user. Second, add cultural terms to the biases and embeddings of an SVD++-style matrix factorization. Results come from one public dataset.

Questions this case answers
  • How to recommend items to a brand-new user who has only a coarse location and no ratings, without using personal data
  • How to add a country-level cultural prior to matrix factorization to predict quality metrics for a user-item pair with no history
  • How to recommend services to a brand-new user when you know only their location
  • How to handle cold start in a recommender without collecting personal data

Tools and methods

  • Hofstede cultural distance — one number per country built from six cultural dimensions
  • Heuristic neighbor selection — three-stage choice of culturally close users and services
  • SVD++-style matrix factorization — biases and embeddings extended with cultural terms
  • WS-Dream dataset — QoS matrices (response time, throughput) with user and service countries
  • MAE and RMSE — offline accuracy metrics
Continue reading — it's free

Google or a one-time email code. No password.