[{"data":1,"prerenderedAt":259},["ShallowReactive",2],{"blog-post-/en/blog/machine-learning-talk-unapec":3},{"id":4,"title":5,"author":6,"body":13,"date":248,"description":249,"extension":250,"image":251,"meta":252,"minRead":253,"navigation":254,"path":255,"seo":256,"stem":257,"__hash__":258},"blog_en/en/blog/machine-learning-talk-unapec.md","Can a Machine Predict Who Will Quit? Notes from a Machine Learning Talk at UNAPEC",{"name":7,"description":8,"username":9,"twitter":10,"avatar":11},"Víctor Garcés","Full-stack developer & content creator","YTvictorworld","https://twitter.com/YTvictorworld",{"src":12,"alt":7},"https://mir-s3-cdn-cf.behance.net/project_modules/disp/49f056206375265.68b77ee9dd6a7.jpg",{"type":14,"value":15,"toc":240},"minimark",[16,35,42,47,66,78,85,89,100,115,126,152,159,166,170,189,196,199,203,232],[17,18,19,20,24,25,34],"p",{},"On July 15 I attended a machine learning talk at ",[21,22,23],"strong",{},"UNAPEC"," (Universidad Acción Pro Educación y Cultura, in Santo Domingo) given by ",[26,27,31],"a",{"href":28,"rel":29},"https://www.linkedin.com/in/ariel-gonzalez-batista-b619a9231/",[30],"nofollow",[21,32,33],{},"Ing. Ariel González Batista",". The room was full of people asking the same two questions I hear everywhere lately: who actually works in this field, and how do I get into it?",[17,36,37,38,41],{},"Instead of answering in the abstract, Ariel built the whole session around one concrete, very human problem: ",[21,39,40],{},"can a machine predict who is going to leave a company?"," Employee retention is expensive to get wrong, so if a model can flag attrition risk early, HR can act before the resignation letter lands. That single question carried us through sixty years of AI history and a full modeling workflow.",[43,44,46],"h2",{"id":45},"a-sixty-second-history-of-ai","A sixty-second history of AI",[17,48,49,50,53,54,57,58,61,62,65],{},"The field of artificial intelligence was born as a discipline around ",[21,51,52],{},"1956",", standing on foundations laid by ",[21,55,56],{},"Alan Turing",". Between the 50s and 60s came the ",[21,59,60],{},"perceptron"," and, later, the idea of stacking them into the ",[21,63,64],{},"multi-layer perceptron",": the ancestor of today's neural networks.",[17,67,68,69,72,73,77],{},"Fast forward to now and the dominant topic is the ",[21,70,71],{},"LLM",". Ariel's reading homework for the room was the paper that made them possible: ",[74,75,76],"em",{},"Attention Is All You Need",". It's on my list.",[17,79,80,81,84],{},"One distinction from that history stuck with me because it explains so much of how these systems feel to use: ",[21,82,83],{},"deterministic vs. probabilistic",". Classical software returns the same answer every time. Machine learning returns a probability, and everything about how you evaluate and deploy it follows from that difference.",[43,86,88],{"id":87},"the-titanic-hypotheses-and-smoke-tests","The Titanic, hypotheses and smoke tests",[17,90,91,92,95,96,99],{},"To make the workflow tangible, Ariel used the classic ",[21,93,94],{},"Kaggle Titanic dataset",": predict who survives the sinking. Everyone knows the historical answer, women and children first, so the interesting lesson is what that knowledge becomes in a model: ",[21,97,98],{},"feature selection",". If survival depended on being a woman or a child, then age and sex are the columns that matter. Look at the data you have and find what is actually important in it.",[17,101,102,103,106,107,110,111,114],{},"I also liked that he took a second to break down the word ",[74,104,105],{},"hypothesis",": from ",[74,108,109],{},"hypo"," (under) and ",[74,112,113],{},"thesis"," (to place). A hypothesis is literally \"what you put underneath\" the work that follows. Everything in the process rests on it.",[17,116,117,118,121,122,125],{},"Then came the part that felt closest to my day job as a developer: before celebrating any model, you need a ",[21,119,120],{},"benchmark"," and a ",[21,123,124],{},"smoke test",".",[127,128,129,136,146],"ul",{},[130,131,132,135],"li",{},[21,133,134],{},"Method 1: the dumb baseline."," About 62% of the passengers died, so a \"model\" that always predicts death is right 62% of the time. Any real model has to beat that number to justify its existence.",[130,137,138,141,142,145],{},[21,139,140],{},"Method 2: a real model."," A ",[21,143,144],{},"random forest",", trained on the features that matter.",[130,147,148,151],{},[21,149,150],{},"Method 3: interpretation."," What is the model actually looking at? If you can't answer that, you don't know what you built.",[17,153,154,155,158],{},"That baseline idea translates directly to engineering: measure the trivial solution first, or you will never know if the sophisticated one is worth anything. As the saying Ariel quoted goes: ",[74,156,157],{},"\"when all you have is a hammer, everything looks like a nail.\""," The baseline is how you check whether you even needed the hammer.",[17,160,161,162,165],{},"All of this, from regression to random forests, lives in ",[21,163,164],{},"scikit-learn",", the Python library that has become the standard entry point to the field.",[43,167,169],{"id":168},"survival-analysis-or-predicting-the-when","Survival analysis, or predicting the \"when\"",[17,171,172,173,176,177,180,181,184,185,188],{},"Back to the employee problem. Predicting ",[74,174,175],{},"whether"," someone leaves is a classification question. Predicting ",[74,178,179],{},"when"," is harder, and that's where the talk introduced ",[21,182,183],{},"survival analysis"," and the ",[21,186,187],{},"Random Survival Forest",": time-to-event models that follow each step of a process and assign it a probability.",[17,190,191,192,195],{},"This is where I raised my hand. I asked how these prediction models interpret ",[21,193,194],{},"censored data",": the employees who haven't left yet, whose story the dataset only knows halfway. The answer connected it back to time-to-event modeling, where those incomplete observations aren't discarded but incorporated, since \"hasn't happened yet\" is itself information.",[17,197,198],{},"My second question was about dimensionality: whether you can generate any amount of data based on the probabilities you want to measure. For scale, the Titanic exercise we discussed worked with a dimensionality of 63 columns. Real problems get wide fast, and knowing which of those columns carry signal is most of the job.",[43,200,202],{"id":201},"what-im-taking-home","What I'm taking home",[127,204,205,211,217,223],{},[130,206,207,210],{},[21,208,209],{},"Baselines before brilliance."," Always measure the trivial predictor first.",[130,212,213,216],{},[21,214,215],{},"Features are the product of domain knowledge."," \"Women and children first\" is history; age and sex as inputs is engineering.",[130,218,219,222],{},[21,220,221],{},"Probabilistic thinking is a different discipline."," Evaluating a model is nothing like testing a function.",[130,224,225,228,229,231],{},[21,226,227],{},"Reading list:"," ",[74,230,76],{},", and the scikit-learn docs.",[17,233,234,235,239],{},"Thanks to Ing. Ariel González Batista and UNAPEC for the session. If you work in ML and think I got something wrong here, or you want to point me to what I should study next, ",[26,236,238],{"href":237},"mailto:victorericksongv@gmail.com","write me",". I'm clearly not done asking questions.",{"title":241,"searchDepth":242,"depth":242,"links":243},"",2,[244,245,246,247],{"id":45,"depth":242,"text":46},{"id":87,"depth":242,"text":88},{"id":168,"depth":242,"text":169},{"id":201,"depth":242,"text":202},"2026-07-15T00:00:00.000Z","I attended a machine learning talk by Ing. Ariel González Batista at UNAPEC. From the Titanic dataset to survival analysis, here are the ideas that stuck with me, and the questions I left asking.","md","/images/blog/machine-learning-talk-unapec.jpg",{},6,true,"/en/blog/machine-learning-talk-unapec",{"title":5,"description":249},"en/blog/machine-learning-talk-unapec","_mi_4dmfyjL5IpYOf636fvzPtuQJpnlgND8_c8G90Iw",1784316712627]