**Our Take: Clustering your furniture customers doesn't require a PhD in dimensionality reduction, it requires a clear question and a willingness to start simple.**
The Reddit user who asked about unsupervised clustering for a furniture and decoration business is asking the right questions, but they may be overcomplicating the answer. K-means is a workhorse, not a magic trick, and it works best when you feed it variables that actually separate your customers into meaningful groups. The temptation to throw everything into a PCA and hope for clarity is understandable, but it often produces a solution that is mathematically tidy and commercially useless.
Start with the business question, not the algorithm. What do you want to know about your customers? Purchase frequency, average order value, product category preferences (sofas vs. lamps vs. wall art), and channel preference (online vs. in-store). Those four variables will already tell you more about your furniture buyers than a dozen engineered features. For categorical variables like product category or payment method, use one-hot encoding, but keep the cardinality low. If you have 50 subcategories, group them into five meaningful families before encoding. K-means struggles with high-dimensional sparse data, so fewer, well-chosen categories are better than many noisy ones.
Variable selection is not a technical problem, it is a judgment call. If you cannot justify why a variable should separate customer groups, remove it. Do not use PCA as a substitute for thinking. PCA can help visualize clusters or reduce noise when you have dozens of correlated variables, but it destroys interpretability. A furniture business owner needs to know that "Cluster A buys modern decor monthly online" not "Cluster A has a PCA component score of 2.3." Run k-means on your selected variables, inspect the cluster profiles, and iterate. If the clusters look like random noise, you either chose the wrong variables or the wrong number of clusters.
The practical takeaway: spend 80% of your effort on variable selection and business interpretation, and 20% on algorithm tuning. Your furniture customers are not a math problem, they are people who bought a dining table last month and might buy a rug next quarter if you show them the right one. PCA will not tell you that. A well-chosen set of purchase behavior variables will.