EM Algorithm for a Two-Component Gaussian Mixture: Log-Likelihood Ascent and Soft Responsibilities
Watch the expectation-maximisation algorithm fit two hidden Gaussian clusters to unlabelled points. Each iteration softly assigns points to components and re-estimates weights, means and spreads while the log-likelihood only ever goes up. Change the separation, weights, spread and sample size and see how overlap slows convergence and biases the estimates. A Blender-built plot table with score-equation and monotonicity checks built in.
Recommendation systems don't just reflect what users like — the act of recommending changes what gets seen, which changes what gets interacted with, which changes what gets recommended next. This simulator makes that closed loop visible: 72 catalogue items sit on a 3D stage as bars whose height and colour track their cumulative exposure, and every round a collaborative-filtering-style rule picks K items to "recommend" with probability proportional to a tunable power of their exposure so far. Push the popularity bias α up and watch a handful of bars pull away into a spike while most of the catalogue stalls at its starting seed; raise the exploration rate ε and watch the distribution stay flat instead. Live Gini coefficient, top-10% exposure share and long-tail percentage quantify exactly how concentrated the catalogue has become, the same metrics used to audit real production recommenders for popularity bias.
Watch the expectation-maximisation algorithm fit two hidden Gaussian clusters to unlabelled points. Each iteration softly assigns points to components and re-estimates weights, means and spreads while the log-likelihood only ever goes up. Change the separation, weights, spread and sample size and see how overlap slows convergence and biases the estimates. A Blender-built plot table with score-equation and monotonicity checks built in.
3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install