Source-linked AI summary
Factorized Point Process Intensities: A Spatial Analysis of Professional Basketball
Andrew Miller, Luke Bornn, Ryan Adams, Kirk Goldsberry
TL;DR
Unadjusted field-goal percentage omits spatial variation in shooting circumstances, while player comparisons often rely on imprecise intuition. The paper models shot locations as related point processes and decomposes their intensity surfaces into low-dimensional features. The resulting representation summarizes shooting habits and outperforms independently fit LGCPs on held-out prediction, while supporting spatial estimates of shooting efficiency.
Problem
Unadjusted field-goal percentage is spatially uninformed and washes away variation due to shooting circumstances.
Source: §6 · p. 9Method
The paper models player shot attempts as related point processes and applies an unsupervised matrix decomposition to their intensity surfaces.
Source: §1 · p. 1Results
The NMF decomposition achieves superior predictive performance over the original independent LGCPs on held-out test data, while providing a compressed representation and interpretable basis.
Source: §5 · p. 9Takeaways & Limitations
The representation provides a low-dimensional summary of shooting habits with an intuitive basis corresponding to recognizable shot types, and supports spatial estimates of shooting efficiency.
Source: §1 · p. 2, §7 · p. 11The treatment does not jointly model spatial field-goal percentage and shot intensity, so it misses their correlation; modeling that effect may improve characterization of players’ habits and ability.
Source: §7 · pp. 11–12
Abstract
from arXiv · showhide
We develop a machine learning approach to represent and analyze the underlying spatial structure that governs shot selection among professional basketball players in the NBA. Typically, NBA players are discussed and compared in an heuristic, imprecise manner that relies on unmeasured intuitions about player behavior. This makes it difficult to draw comparisons between players and make accurate player specific predictions. Modeling shot attempt data as a point process, we create a low dimensional representation of offensive player types in the NBA. Using non-negative matrix factorization (NMF), an unsupervised dimensionality reduction technique, we show that a low-rank spatial decomposition summarizes the shooting habits of NBA players. The spatial representations discovered by the algorithm correspond to intuitive descriptions of NBA player types, and can be used to model other spatial effects, such as shooting accuracy.
1. Introduction.
The paper develops an unsupervised low-dimensional representation of related point-process intensities to summarize NBA players’ shooting habits. It relates this approach to prior basketball and neural-data models while emphasizing interpretable player features.
1. Introduction.
- The method decomposes Poisson-process intensity surfaces to represent related point processes in a low-dimensional, interpretable feature space.The components quantitatively summarize offensive tendencies and characterize player types better than traditional categories such as position.Source: §1 · p. 1
- The representation can support personnel selection for diverse offensive tendencies and spatial estimates of players’ field-goal percentage.The paper uses a latent-variable model to visualize field-goal percentage across court locations.Source: §1 · pp. 1–2
1.1. Related Work.
- Prior NBA analyses either predict game scores without spatial or individual-player modeling, or focus on defensive effects near the basket for a subset of players.The paper contrasts its scope with Adams et al. (2010) and Goldsberry and Weiss (2012).Source: §1.1 · p. 2
- A related neural-data approach discovers latent trajectories, whereas this model’s positivity constraint changes the latent functions’ behavior and interpretation.The comparison is with Yu et al. (2009).Source: §1.1 · p. 2
2. Background.
The background introduces Gaussian and Poisson processes, log-Gaussian Cox processes, and non-negative matrix factorization as ingredients for modeling spatial shot patterns. Together, these concepts motivate smooth intensity surfaces and a nonnegative low-rank decomposition.
2.1. Gaussian Processes.
- Gaussian processes model functions probabilistically, with spatial covariance encoding prior beliefs such as smoothness; the paper uses this to assume shooting habits vary smoothly across the court.The squared exponential covariance example assumes infinitely differentiable functions.Source: §2.1 · p. 2
2.2. Poisson Processes.
- A Poisson process assigns Poisson-distributed counts to spatial sets, and independent processes can be superposed.Because intensity functions can be scaled by nonnegative factors, the superposition property motivates decomposing a global process into weighted subprocesses shared between players.Source: §2.2 · pp. 2–3
2.3. Log-Gaussian Cox Processes.
- A log-Gaussian Cox process is a doubly stochastic Poisson process whose spatially varying intensity is modeled as an exponentiated Gaussian process.Its two sources of randomness are the random function and the point process conditional on its intensity.Source: §2.3 · p. 3
2.4. Non-Negative Matrix Factorization.
- Non-negative matrix factorization approximates a matrix with low-rank factors, using nonnegative weights to scale and combine basis vectors.The positivity constraint disallows negative bases from canceling positive ones, tending toward sparser, more interpretable bases than less constrained methods such as PCA.Source: §2.4 · pp. 3–4
- The choice of reconstruction loss changes which regions the factorization emphasizes: KL divergence penalizes large ratios, while Frobenius loss favors larger-intensity regions.Under KL loss, low-intensity regions can contribute more when their original-to-reconstructed ratio is large.Source: §2.4 · p. 4
3. Data.
The study uses made and missed shot locations from roughly half of 2012–2013 NBA regular-season games to represent player shooting on the offensive half court. Its exploratory spatial correlations show long-range, non-stationary patterns, especially among three-point shooters.
3. Data.
- About 78,000 shots from 335 players, after excluding players with fewer than 50 attempts, form the dataset.The shots come from roughly half of the 2012–2013 regular-season games.Source: §3 · p. 4
- Player shooting is modeled as a point process on a 35 ft by 50 ft offensive half court, with made and missed attempts distinguished by outcome.Figure 1 shows the original shot locations, discretized counts, LGCP surfaces, and NMF reconstructions for two players at K = 10.Source: §3 · p. 4, Figure 1 · p. 5
3.1. A Note on Non-Stationarity.
- Long-range shot-count correlations follow non-stationary patterns, particularly for three-point shooters, motivating NMF to learn global spatial patterns beyond a stationary LGCP’s locally smooth surface.The correlation analysis uses counts from tiles of a discretized court.Source: §3.1 · p. 5
4. Proposed Approach.
The approach estimates spatial intensity surfaces for players, then uses non-negative matrix factorization to find shared shot-pattern bases and player-specific weights. The authors compare this representation with alternative approaches to modeling the spatial structure.
4. Proposed Approach.
- The pipeline discretizes shot locations, estimates each player’s LGCP intensity surface, normalizes it, and factorizes the resulting matrix into nonnegative bases and player loadings.The court is divided into one-square-foot tiles for computational tractability; the authors expect this resolution to capture interesting spatial variation.Source: §4 · pp. 5–6
4.2. NMF Optimization.
- Nonnegative bases act as sub-intensities, and each player’s linear weights summarize the amount of each shot type in their shooting habits.The table groups normalized player weights into close-range, mid-range, and three-point shot bases.Source: §4 · pp. 5–6, Table 1 · p. 7
4.1. Fitting the LGCPs.
- A per-player tile-count Poisson likelihood and Gaussian-process prior define the latent-field posterior; elliptical slice sampling estimates it, and posterior mean intensities are retained.The Gaussian-process covariance encodes spatial structure, and its bias term parameterizes the mean Poisson rate.Source: §4.1 · p. 6
4.2. NMF Optimization.
- The authors compare KL and Frobenius loss functions to examine how the resulting basis vectors differ.The visual comparison presents LGCP-NMF with both losses, NMF on raw counts, and PCA on LGCP intensity functions.Source: §4.2 · p. 6, Figure 4 · p. 8
- With K = 10, the learned bases include corner, wing, and top-of-key three-point shot types, each represented by a normalized intensity surface.Conditioned on a shot type, its surface acts as a density over shot locations.Source: Figure 3 · p. 7
4.3. Alternative Approaches.
- LGCP-NMF learns natural basketball-court symmetries without user guidance, while raw-count NMF yields poorer, degenerate bases and PCA produces mixed-sign, visually unintelligible bases.Raw-count factorization is less numerically stable and requires jitter in the intensity matrix for convergence.Source: §4.3 · p. 7, Figure 4 · p. 8
5. Results.
LGCP-NMF represents NBA shooting with interpretable shot-type bases and player-specific weights. Its low-rank reconstruction improves held-out predictive performance over independent LGCPs.
5. Results.
- The learned basis vectors correspond to visually interpretable shot types, including corner, wing, and top-of-the-key threes.The displayed examples use K = 10 and KL-based loss.Source: §5 · p. 8
- Player-specific basis weights quantify how much each player takes each shot type, summarizing offensive habits.Table 1 compares normalized weights for selected players.Source: §5 · p. 8
- KL-based NMF yields more spatially diverse bases than Frobenius-based NMF, which concentrates on the high-intensity region near the basket.The comparison is shown in Figure 4(a–b).Source: §5 · pp. 8–9
- Direct count-matrix NMF produces sparse, spatially unstructured bases, while PCA yields mixed-sign bases that render them uninterpretable as intensity functions.NMF avoids PCA’s cancellation of positive and negative bases representing the same feature.Source: §5 · p. 9
- LGCP-NMF achieves superior held-out predictive performance to independent LGCPs while providing a compressed, interpretable representation.Figure 5 evaluates average player test log likelihood, holding out 10% of each player’s shots per fold.Source: §5 · p. 9, Figure 5 · p. 9
6. From Shooting Frequency to Efficiency.
The paper extends shot-type representations to spatial shooting efficiency, addressing the loss of location-specific information in unadjusted field goal percentage. A latent-variable model combines shot-type efficiencies with player-specific shot patterns, and the resulting surfaces support regional player comparisons.
6. From Shooting Frequency to Efficiency.
- Unadjusted field goal percentage ignores shooting circumstances, while the proposed spatial estimates describe efficiency across locations on the court.The model aims to estimate each player’s probability of making a shot at each point in the offensive half court.Source: §6 · p. 9
6.1. Latent variable model.
- The latent-variable model combines shot-type field goal percentages using location-dependent shot-type probabilities, with player weights adjusted for basis mass.Shot type is latent; observed shot location is used to compute the predictive distribution.Source: §6 · p. 9, §6.1 · p. 10
- Hierarchical priors share information across players within each shot type and shrink estimates for players with small sample sizes toward the global mean.Posterior samples of the model parameters are drawn using Gibbs sampling, elliptical slice sampling for β given σ² and outcomes, and conjugacy for σ² given β and outcomes.Source: §6.1 · p. 10, §6.2 · p. 10
6.3. Results.
- The fitted surfaces suggest Irving and Novak are below average on the baseline jump shot, while Curry is above average on corner threes.Global mean efficiency and its posterior uncertainty are shown alongside selected player-specific spatial efficiency surfaces.Source: §6.3 · pp. 10–11, Figure 6 · p. 11
7. Discussion.
The discussion presents constrained matrix decomposition of independently fit intensity surfaces as a way to summarize related point processes and NBA shooting habits. It also identifies jointly modeling shot frequency and spatial efficiency as a direction that could better capture their correlation.
7. Discussion.
- Constrained matrix decomposition of independently fit intensity surfaces provides a low-dimensional summary of related point processes and interpretable shot-type bases.The discussion describes the representation as an intuitive summary of shooting habits and uses it to form spatial shooting-efficiency estimates.Source: §7 · p. 11
- A future fully generative model could jointly model point processes and intensity decomposition, potentially allowing richer representations that include defensive effects.The proposed direction draws on the relationship between KL-based NMF and probabilistic latent-variable models.Source: §7 · p. 11
- The current treatment misses the correlation between player skill and shot locations; jointly modeling spatial efficiency and intensity may better characterize habits and ability.The discussion notes the intuition that players take more shots from locations where they are more accurate.Source: §7 · pp. 11–12