← Model library

Compare two model types

Pick two kinds of model to see what each one does, the data it needs and how to start, side by side. Not sure which kinds to look at? Start from which model you need.

First model type Second model type

Two model types compared
	Tabular classification	Number prediction (regression)	Time-series forecasting	Text classification	Image classification	Object detection	LLM assistant (RAG chatbot)	Recommendation system	Anomaly detection	Speech and audio
What it does	Predict a category (yes/no, A/B/C) from rows of a spreadsheet or database.	Estimate a numeric value such as a price, a duration or a score.	Predict future values from history: demand, traffic, sales, load.	Sort text into categories: sentiment, topic, intent, spam, priority.	Assign a label to a whole image: defect/ok, species, product type, diagnosis.	Find and locate objects in images or video with bounding boxes; count and track them.	A chatbot or assistant that answers questions using your own documents and data.	Suggest relevant products, content or people to each user.	Spot unusual events: failing machines, suspicious logins, odd transactions.	Transcribe speech, classify sounds or build voice interfaces.
Typical projects	Predict which customers will churn Flag fraudulent transactions Approve or reject loan applications	Estimate house prices from features Predict delivery time for an order Estimate insurance claim cost	Forecast weekly sales per store Predict website traffic next month Forecast energy demand per hour	Detect sentiment of product reviews Route support tickets to the right team Filter spam emails	Detect defective products on a production line Identify plant diseases from leaf photos Classify X-ray images	Count people entering a store from CCTV Detect helmets on construction workers Find damaged areas on cars	Chatbot that answers questions from our PDF manuals Internal assistant over company wiki Customer support bot using our FAQ	Recommend products based on purchase history “You might also like” for articles Suggest courses to learners	Detect machine failures from sensor data Find unusual login activity Spot abnormal spikes in server metrics	Transcribe customer calls and summarize them Voice commands for a mobile app Detect machine sounds that indicate faults
The data you need	Export one row per entity (customer, transaction) with the outcome you want to predict as a column.	Collect one row per item with the true value you want to predict.	You need a timestamp column and the value, ideally 2+ full seasonal cycles (e.g. 2 years for yearly patterns).	Collect real text examples from the place the model will run (tickets, reviews).	Collect images in the same conditions as production (lighting, angle, camera).	Collect frames covering different lighting, distances and occlusions.	Gather the documents the bot should know (PDF, HTML, Notion, Confluence…).	Interaction logs are the key: user_id, item_id, timestamp, event (view/click/buy).	Collect mostly “normal” history — anomalies are rare by definition.	Collect real recordings: same microphones, accents, background noise as production.
Labels	Labels usually already exist in your systems (for example, a “canceled subscription” flag). Join them from your CRM / billing database.	The target is usually a historical number (sold price, actual delivery minutes). Pull it from your transactional database.	No manual labeling needed — the future values in your history are the labels.	Use Label Studio or Argilla. Or bootstrap: let an LLM pre-label and have humans correct it.	Folder-per-class is enough. For larger sets use Label Studio, CVAT or Roboflow.	Draw boxes in CVAT, Label Studio or Roboflow. Use a pretrained model to pre-annotate, then correct its labels.	No training labels needed. Your golden Q&A set is the evaluation “label”.	Implicit feedback (clicks, purchases) acts as labels — no manual labeling.	Usually unlabeled. Label a small set of confirmed incidents for evaluation; domain experts review alerts.	Correct auto-generated transcripts rather than typing from scratch (Label Studio supports audio).
GPU needed?	Usually not: a regular CPU is enough.	Usually not: a regular CPU is enough.	Usually not: a regular CPU is enough.	Often, to train or fine-tune your own. Starting with an API needs none.	Often, to train or fine-tune your own. Starting with an API needs none.	Often, to train or fine-tune your own. Starting with an API needs none.	Usually not: a regular CPU is enough.	Usually not: a regular CPU is enough.	Usually not: a regular CPU is enough.	Often, to train or fine-tune your own. Starting with an API needs none.
How it’s measured	F1 score and ROC-AUC (use precision/recall if one kind of mistake is costlier)	MAE (easy to explain: “off by $X on average”) and RMSE (punishes big misses)	MAPE or MAE measured with backtesting over several past periods	F1 score per class (macro-F1 when classes are imbalanced)	Accuracy and per-class recall (for defect detection, missed defects matter most)	mAP (mean Average Precision) at IoU 0.5, plus FPS for video	Answer correctness and groundedness on a golden Q&A set; latency and cost per answer	Recall@K / NDCG offline; click-through and conversion in an A/B test online	Precision at a fixed alert budget (e.g. top 50 alerts/day) and recall on known incidents	WER (Word Error Rate) for transcription; accuracy/F1 for sound classification
Easiest way to start	AutoML (e.g. AutoGluon / Google Cloud AutoML / SageMaker Autopilot)	AutoML regression (AutoGluon / Google Cloud AutoML)	Pretrained forecasting model (Chronos / TimesFM) or Prophet	LLM zero-shot / few-shot classification via API	No-code / AutoML vision (Teachable Machine, Roboflow, Google Cloud AutoML)	Pretrained YOLO / open-vocabulary detector (YOLO-World, Grounding DINO)	Hosted LLM API (Claude) + managed RAG / long context	Popularity + content-based similarity (embeddings of item descriptions)	Statistical rules + Isolation Forest	Cloud speech-to-text API or Whisper (pretrained)
What most teams use	Gradient-boosted trees (XGBoost / LightGBM)	Gradient-boosted trees (LightGBM / XGBoost)	StatsForecast (AutoETS/AutoARIMA) + LightGBM with lag features	Fine-tuned small transformer (DistilBERT / ModernBERT) or SetFit	Fine-tuned pretrained CNN/ViT (ConvNeXt, EfficientNet) via timm	Fine-tuned YOLO (Ultralytics) on your labeled images	RAG: hosted LLM + embeddings + vector DB	Collaborative filtering (implicit ALS / LightFM) + content features	Isolation Forest / ECOD ensembles via PyOD with rolling features	Whisper (faster-whisper) self-hosted + LLM for summaries
Watch out for	Data leakage — a feature that secretly contains the answer	Extrapolation — models can’t predict well outside the range they saw	Random train/test split leaks the future	Inconsistent labels between annotators	Model learns the background, not the object	Tiny objects need higher input resolution	Hallucination when retrieval misses — instruct the model to say “I don’t know”	Random splits leak future behavior	Alert fatigue from too many false positives	Testing on clean audio but deploying in noisy environments
Guides	Predict customer churn from your order history Detect fraudulent transactions Build your first machine learning model in Google Colab	Predict delivery times Predict house prices Predict customer lifetime value	Forecast daily sales for each store Forecast call volume and plan staffing	Build a spam filter, start to finish Sort customer reviews by sentiment Route support tickets to the right team Label data for machine learning	Find defective products in photos Sort product photos into categories	None yet. The model page covers the basics.	Build a chatbot that answers questions from your PDFs Build a support chatbot from your help center	Recommend products your customers will like	Spot failing machines from sensor data Flag unusual expense claims	None yet. The model page covers the basics.
Next steps	Plan one Read more	Plan one Read more	Plan one Read more	Plan one Read more	Plan one Read more	Plan one Read more	Plan one Read more	Plan one Read more	Plan one Read more	Plan one Read more