Title: Analysis of Decision Support System of Basketball Sports Competition Based on Machine Learning and IoT
Author: Muren Bao
Institution: College of Physical Education, Inner Mongolia University; Scientific Reports
Publication Date: June 22, 2026
The Problem: One Possession, Too Many Disconnected Reads
Traditional basketball tactical analysis still leans on manual observation which is subjective and difficult to scale across a full offensive and defensive possession. Prior machine learning work has mostly tackled single slices of the game with trajectory tracking, shot detection, or defensive pattern classification in isolation.
That fragmentation makes it hard to connect how a particular formation produces a good shot, leaving coaches with scattered data points instead of one coherent read on why a possession succeeded or failed. IoT sensor data and machine learning models are also rarely integrated end to end, limiting real-time tactical feedback.
Methodology: Four Layers Working Together
Three layer IoT capture (the scouting crew): Six ceiling cameras plus player wearables feeding an edge gateway collect position, motion, and heart rate data in sub-millisecond sync like a scouting crew wired directly into the court.
Modality specific encoding (the translator): Position, kinematics, and physiological signals each get their own encoder before merging, similar to translating box-score stats and video notes into one shared language before comparing them.
Bidirectional LSTM (the film reviewer): The model reads each possession forwards and backwards at once, the way a coach reviewing film weighs both what led into a play and what followed it.
Multi-head attention + joint training (the highlight editor): An attention layer dials up focus on moments that matter most like shot release, pick and roll formation and defensive rotations while classification and shot probability prediction train together so each task sharpens the other.
Why it Matters: Strong Accuracy, With Real Limits
89.7% classification accuracy: Tested on NBA SportVU tracking data from 2018–2023 (42,000 possession samples), the model correctly read offensive-defensive states 89.7% of the time on the test set.
Lowest shot error rate tested: Its shot probability prediction posted a 9.8% MAPE, beating a Bi-LSTM plus attention baseline by a statistically significant margin (p < 0.001 on most metrics, five runs each).
Cracks under exactly the pressure coaches care about: Accuracy drops from roughly 90% to 85% under heavy player occlusion near the basket, and shot error rates jump nearly 25% on possessions inside the final two seconds of the shot clock.
Infrastructure heavy: The system depends on a full six camera, wearable sensor rig and A100-class GPUs This is realistic for an NBA level program, not a plug and play tool for smaller programs without that data pipeline.
ACTIONABLE TAKEAWAYS
Pair classification with shot prediction, not either alone: The joint model beats single task versions on both fronts. Deploying only a state classifier or only a shot model leaves accuracy on the table.
Bucket late clock and occluded possessions separately: Flag possessions under 6 seconds on the shot clock and heavily occluded sequences as their own analytic category rather than blending them into season wide averages, since error rates spike in both.
Audit your camera setup before you trust the data: Before investing in a full tracking rig, confirm your footage can support reliable player identification under traffic near the basket. That's where accuracy erodes most.