Mondrian Forests

Foundations and Beginner Paper Guide

A visual 32-page introduction from decision trees to online updates

Prepared for MSc Data Science study | Based primarily on Lakshminarayanan, Roy and Teh (2014)

How to use these notes

Read in order once; return to pages 12-31 while reading the paper.

The central question

How can a random-forest-style model learn from new observations one at a time without rebuilding every tree from the beginning?

1

Decision trees

We begin with the standard model. One model divides the feature space into regions.

2

Forests

A single tree is unstable. Many random trees combine their predictions to form an ensemble.

3

Online learning

The core problem. New observations arrive sequentially and the model must adapt immediately.

4

Mondrian process

A mathematical tool. A random, time-labelled partition creates extendable trees.

5

Paper mechanics

The specific methodology in the 2014 paper. Training, extension, prediction and label smoothing.

Scope

These notes explain the 2014 classification paper entirely within the context of tree-based methods. Later Mondrian variants expand the idea to regression tasks and alternative online aggregation.

1. Supervised learning in one picture

Features go in; a class label or numerical value comes out.

Observed data

$x = \text{features, } y = \text{answer}$

Learning algorithm

Find useful patterns in the historical data

Fitted model

Stores the learned mathematical rules

Prediction

Estimate $y$ for a completely new $x$

feature 2 feature 1

Classification

The answer is a category. Examples include predicting cat/dog, fraud/not fraud, or class 1/2/3. The 2014 Mondrian paper evaluates classification tasks.

Regression

The answer is a continuous number. Examples include predicting a house price, product demand, or a continuous risk score.

Features define a space

With two features, each observation is a point on a flat plane. A classifier draws boundaries that divide this plane into distinct regions. A decision tree specifically draws axis-aligned boundaries, creating vertical or horizontal cuts across the space.

2. How a decision tree thinks

A sequence of yes/no questions routes each observation to a leaf.

yes no yes no yes no x₁ ≤ 4.5? x₂ ≤ 2.0? x₂ ≤ 6.0? ClassA ClassB ClassB ClassC

Internal node

An internal node contains a logical test, such as $x_1 \le 4.5$. This test routes the observations into two separate groups based on their feature values.

Leaf node

A leaf node represents a terminal region. The tree assigns a prediction to this region based on the labels of the training observations that fall into it.

The selection of splits

Standard trees search through many possible features and cut points. They prefer the cut that produces the purest child groups. Algorithms like CART commonly use Gini impurity to measure this purity. Other trees may use entropy for classification or variance reduction for regression.

Important contrast

Mondrian-tree splits are sampled from geometry. They do not use class labels to choose the split location or dimension. The tree relies on labels entirely after the structure is built to make its final predictions.

3. Split quality: a small worked example

This demonstrates standard-tree behaviour. Mondrian trees select splits differently.

Parent node: 10 observations

Six blue observations and four orange observations are mixed together. A useful split should make the resulting child nodes less mixed.

Candidate A: useful split

child 1
child 2

Candidate B: weak split

child 1
child 2

Gini impurity mathematical definition

For class proportions $p_1, \dots, p_K$, the calculation is $\text{Gini} = 1 - \sum_{k=1}^K (p_k^2)$. The value is $0$ for a perfectly pure node. The value increases when classes are thoroughly mixed within the node.

NodeBlue ($p_1$)Orange ($p_2$)Gini Score
Parent0.600.400.48
A - child 11.000.000.00
A - child 20.200.800.32
B - each child0.600.400.48

Takeaway

Standard trees evaluate exact label distributions to optimise split quality. Mondrian trees trade that label-driven optimisation for random, geometric splits. These random splits possess a special online extension property.

4. From one tree to a forest

Bagging reduces instability by averaging many varied trees.

Bagging workflow

training data
bootstrap 1
bootstrap 2
bootstrap 3

combine tree probabilities / votes

Random Forest

A standard Random Forest uses bootstrap samples and random feature subsets to build independent trees. Splits within those trees are normally chosen using labels and an impurity criterion.

Mondrian Forest

A Mondrian Forest uses independent Mondrian trees. Its split geometry is completely random and supports consistent online extension. It still averages the predictions from multiple trees.

Common principle

Both methods are ensembles. Several imperfect and varied trees combine their individual calculations to produce a steadier, more accurate final prediction.

5. Batch learning and online learning

The difference concerns when data become available and how the model incorporates updates.

Batch learning

collect dataset
train once
deploy
retrain later

Online learning

observe $x_1$
update
observe $x_2$
update
observe $x_3$
update

Prequential evaluation

In a streaming environment, a common testing sequence is test-then-train. The model predicts the label for the incoming point, records the error for evaluation purposes, and then updates its internal structure once the true label is revealed.

Online learning is not automatically concept-drift handling

A model may update online while still treating all old observations as permanently relevant. Drift-aware systems often need mechanisms like forgetting, observation weighting, fixed windows, or explicit change detection. The original Mondrian Forest paper focuses exclusively on efficient online extension. It does not present a complete drift-management system.

Distribution shift

The input appearance changes over time, such as changing lighting or backgrounds in images.

Concept drift

The fundamental relationship between the features and the correct answer changes over time.

06 / THE STREAMING PROBLEM

6. The streaming data problem

A repeated batch workflow may keep historical observations and train a new forest whenever the dataset grows. The cost of each retraining cycle becomes larger as the stream continues.

×

Repeated batch retraining

Historical data are commonly retained or reloaded, and the complete model is fitted again.

Mondrian online updates

The existing trees are extended. The model stores node geometry and label statistics rather than rebuilding every tree.

The chart is a schematic resource comparison. The 2014 paper reports computational complexity and training times rather than RAM traces.

Illustrative memory pressure over time

A slowly growing online model state is compared with a workflow that retains an expanding history for repeated fitting.

7. Trees as geometric partitions

Every leaf corresponds to a rectangular region of feature space.

No splits

One split

Three splits

Axis-aligned split

A rule such as $x_d \le x_i$ cuts one rectangular region into two. Repeating this operation creates the geometric hierarchy represented by the decision tree model.

Name intuition

The rectangular partitions resemble the grid-like compositions associated with the artist Piet Mondrian. The underlying mathematical object that generates these cuts is called a Mondrian process.

Tree view

This perspective is useful for following a prediction sequentially from the root node to a leaf node.

Partition view

This geometric perspective is useful for understanding the spatial limits and behaviour of Mondrian cuts.

8. The Mondrian process: intuitive construction

A random clock decides when a rectangle splits; its side lengths influence the clock rate.

width = $u_1 - l_1$ height = $u_2 - l_2$

Waiting time

The variable $E$ is an exponential waiting time. The rate of this exponential clock is equal to the sum of the region's side lengths. In two dimensions, this rate is simply the width plus the height.

Split direction

The algorithm must choose a dimension to split. It makes this choice with a probability proportional to each side length. Longer dimensions possess a higher probability of being cut.

Split location

After choosing the dimension, the exact cut point is chosen uniformly along that side. The algorithm then repeats this entire random process independently inside the two newly created child rectangles.

Large rectangles tend to split sooner

small box
medium box
Because the exponential rate is the sum of the dimensions, larger dimensions result in a higher rate. A higher rate produces a shorter expected waiting time before the clock triggers a cut.

Technical note

The paper applies this process exclusively to the bounding boxes of observed training points. If the process were applied to an infinite space, it would run forever. Bounding the process ensures the stored tree remains finite.

9. Split time and the lifetime parameter

Each node has a recorded time. The lifetime parameter lambda stops tree growth.

0
$1.2$
$2.7$
$4.0 = \lambda$
$\tau=2.7$
$\tau=2.1$

Small lambda

A small lambda prevents the tree from running its clocks for long. This results in few splits, large leaf regions, and a smoother, simpler partition model.

Large lambda

A large lambda allows the internal clocks to run longer. This generates more splits, smaller leaf regions, and a highly detailed partition of the feature space.

lambda = infinity

The implementation in the 2014 experiments pauses pure leaves rather than stopping at a finite lifetime. The tree only stops expanding when all labels in a block are identical.

Analogy

The lambda parameter plays a role similar to a complexity control such as maximum depth in standard random forests. However, lambda operates continuously in time rather than simply counting the discrete number of nodes.

10. The observed-data bounding box

Mondrian splits are committed only where training data have actually been observed.

new point
Bx: smallest rectangle containing node data

The role of the uncommitted exterior

Outside the grey boundary box, the old tree structure has made no fixed geometric commitment. A new observation arriving in this space can therefore trigger the creation of an earlier split that was never needed previously.

Inside the box

If a new observation arrives inside the boundary box, the algorithm follows the existing split structure down the tree. It then simply updates the extent of the relevant node.

Difference from an ordinary displayed tree

A conventional decision tree diagram often suggests that each rule applies infinitely across the entire parent region. A Mondrian split is strictly tied to the observed-data extent at that node. This uncommitted exterior is the exact property that makes online extension computationally feasible.

Notation from the paper

$B_j$ denotes the full spatial block represented by a node. $B_j^x$ denotes the smallest axis-aligned rectangle that entirely encloses the training observations currently held at that node.

11. Handling new data immediately

When a point lies outside the current observed-data box, the tree measures the extra extent. It may insert a new split above the current node if the sampled split time is earlier.

1. Current observed box

The node records the smallest rectangle that contains the observations it has already seen.

2. Stream arrival

The new point lies outside the known extent, so the current rectangle must expand.

3. Expand and possibly insert

The expanded box contains all points. A sampled earlier split can become a new parent; otherwise the existing node is simply extended.

Important detail: an outside point does not always create a new root. The exponential waiting-time test decides whether a new split comes before the node's current split time.

12. Adding one new observation

The online update either inserts a new parent node or extends an existing node and recurses.

A. Point inside extent

B. Point just outside

C. Point far outside

1
Measure extra extent

The algorithm measures how far $x$ sits beyond the current lower or upper boundaries across every dimension.

2
Sample a possible new split time

The algorithm draws a new waiting time from an exponential distribution. The rate of this new clock equals the total extra extent.

3
Compare times

If the newly sampled time is earlier than the existing node's split time, a new parent node is inserted above the current position.

4
Otherwise

If the new time is later, the algorithm simply expands the bounding box to enclose the new point and recurses down the existing tree path.

The distance effect in practice

A point located further away produces a larger extra extent. This larger extent creates a faster exponential clock, increasing the mathematical probability that a new split will be inserted above the existing node.

13. How a new parent is inserted

A distant new point can create a split above an older split while the old subtree remains intact.

Before the new point arrives

Old root $x_2 \le 2.1$ Leaf A Leaf B
new distant point

After an earlier split is sampled

New parent $x_1 \le 5.5$ Old root preserved $x_2 \le 2.1$ New leaf new point Leaf A Leaf B

The old work is kept

The old root and both old leaves remain as one complete subtree.

The new split is earlier

Its sampled split time occurs before the old root split time, so it becomes the parent.

One new branch is added

The arriving point occupies the sibling leaf on the other side of the new split.

14. The key property: projectivity

A tree restricted to a subset of points follows the identical probability law as a tree built directly on those points.

Batch route

  • The system receives all $N+1$ points simultaneously.
  • The algorithm runs the standard Mondrian procedure and samples one complete tree.
  • This yields the final probability distribution $MT(\lambda, D_{1:N+1})$.
Same
Distribution

Online route

  • The algorithm builds a tree on the initial sequence of $N$ points.
  • A new observation, point $N+1$, arrives later.
  • The algorithm samples a conditional extension using the distance mathematics.
  • This sequence produces the identical final probability distribution.

What "same distribution" means

The individual random trees generated by the two routes need not be visually identical at the node level. The property simply means the online update procedure samples from the exact same mathematical probability law as a fresh batch construction on the enlarged dataset.

The practical benefit of projectivity

  • Random decisions made early in the stream remain completely compatible with data that arrives later.
  • The final structure of the forest does not depend on the specific order in which the observations arrive.
  • Updating an existing tree requires following a single path from the root to a leaf, bypassing the enormous computational cost of rebuilding the entire forest.

The core theoretical contribution

This strict batch-online distributional agreement is the primary theoretical feature distinguishing Mondrian Forests from earlier online random-forest schemes, which often relied on approximations or heuristic split-decisions.

15. Labels and prediction smoothing

Tree geometry is generated randomly; label information enters through class counts and a mathematical hierarchy.

Root Prediction [0.50, 0.50] Parent Prediction [0.62, 0.38] Leafcounts 8 blue, 1 orange

The necessity of smoothing

Because the Mondrian splits are random, a leaf may ultimately contain very few observations. Relying solely on the raw frequency of a tiny sample creates extreme, unreliable probabilities. To counteract this, the paper implements a hierarchical Bayesian label model (specifically using an interpolated Kneser-Ney approximation). This structure forces a node to borrow statistical strength from its parent.

Near parent in split time

If a child node is created very soon after its parent, the algorithm forces the child's prediction to remain highly similar to the parent's established prediction.

Far from parent in split time

If the child node represents a split occurring significantly later in time, the node places much greater weight on the specific observations within its own local region.

16. Predicting a point outside the observed region

The method calculates an average over the locations where the test point could branch away.

1 node 1 2 node 2 3 node 3 possible branch possible branch possible branch

Probability of separating at a node

The algorithm computes the mathematical probability that the test point branches off. This calculation requires the test point's specific distance outside the original bounding box and the continuous exponential rate associated with that node.

Calculating the final prediction

The algorithm adds the class-probability contribution from every theoretical branch point. It weights these contributions by the probability of reaching and separating at that specific location. If the mathematics dictate no separation occurs, the system simply uses the reached leaf prediction.

The practical consequence

Predictions systematically become less certain as a test point moves further away from the regions actively supported by training data. The model acknowledges uncertainty rather than blindly extending the nearest leaf rule into infinity.

17. The paper's algorithms as one workflow

The original pseudocode detailed in the paper separates the logic into four distinct computational jobs.

1. Build

SampleMondrianTree

The foundational step. The algorithm recursively draws split times, dimensions, and locations to construct the initial unlabelled geometric structure.

2. Update

ExtendMondrianTree

The online adaptation step. The algorithm inserts a parent node or expands an existing node boundary when a new labelled point arrives from the data stream.

3. Count labels

Posterior counts

The statistical step. The system updates the class frequencies along the affected path to prepare the hierarchical Bayesian smoothing approximations.

4. Predict

Predict

The evaluation step. The algorithm traverses the updated tree and calculates the weighted average of possible test-point extensions.

Complexity intuition

Assuming the tree remains reasonably balanced, a single update or prediction follows a path whose length grows proportionally to $\log N$. The original paper notes that the total processing cost scales as $\mathcal{O}(N \log N)$. In contrast, repeating a full batch retraining after every single data arrival produces a heavy computational burden scaling as $\mathcal{O}(N^2 \log N)$.

18. Four related tree approaches compared

The methods may all process trees, but their statistical ideas and update rules differ.

Method Data mode How splits are chosen How predictions combine New-data treatment Exact Mondrian batch-online property
Batch Random Forest Batch Labels guide impurity reduction; feature subsets add randomness Model combination across many trees Usually rebuild or refit on an enlarged dataset No
Mondrian Forest Online or batch Random geometric cuts independent of labels Model combination across independent Mondrian trees Extend bounding boxes and sometimes insert an earlier parent Yes
ORF-Saffari Online Candidate split scores at leaves use labels and confidence rules Model combination across online trees Update leaf statistics; commit a selected split when rules are met No
Dynamic trees Online Bayesian method Posterior tree proposals use feature and label likelihood information Bayesian model averaging across particles Update a particle approximation to the tree posterior No

Model combination

Random Forests and Mondrian Forests keep a diverse collection of trees and average their outputs.

Bayesian model averaging

Dynamic trees weight particles as an approximation to uncertainty about the underlying tree model.

Drift remains separate

None of these labels by itself means that old concepts will be forgotten when the data-generating relationship changes.

19. Computational scaling: update versus repeated retraining

The paper compares cumulative work under balanced-tree assumptions.

Growth of cumulative computational work

Logarithmic vertical scale; curves come from the stated complexity formulas rather than measured seconds.

10 20 50 100 200 500 1000 OBSERVATIONS PROCESSED LOG CUMULATIVE WORK Mondrian: $N\log N$ retrained batch: $N^2\log N$

Mondrian updates

Each arriving point usually follows one root-to-leaf path. Under a balanced-tree assumption, that path has length proportional to $\log N$.

Repeated batch retraining

The entire forest is fitted again after each arrival. Work accumulates across successively larger datasets.

Reported experimental message
10×+

The authors report at least an order-of-magnitude speed advantage against repeated batch retraining and the compared earlier online forest on several datasets.

Interpretation boundary

The exact runtime depends on implementation, hardware, tree balance, feature count and the chosen comparison procedure. The chart above explains asymptotic growth only.

20. What the 2014 experiments tested

The authors tested predictive accuracy as more data arrive sequentially, and accuracy against total training time.

Accuracy versus fraction of data

FRACTION SEEN TEST ACCURACY

Accuracy versus training time

TIME (LOG SCALE) TEST ACCURACY
Mondrian Forest
batch forest
earlier online RF

Results on experimental datasets

On standard datasets like USPS, Satimages, and Letter, Mondrian Forests achieved accuracy very close to the batch forest variants. The algorithm exceeded the accuracy of the compared online forest, while processing the stream much faster than repeated batch retraining.

The algorithm performs poorly on the DNA dataset. The random splitting mechanism does not use class labels, meaning the tree wastes splits on irrelevant features instead of isolating useful signals.

Chart caution

The plots above are illustrative sketches intended to show the general convergence trends. They are not digitised measurements. Consult Figure 3 of the original paper for the precise reported curves and experimental parameters.

21 / PERFORMANCE

21. The speed and accuracy comparison

The paper reports accuracy close to strong batch forests on several datasets, while training is at least an order of magnitude faster than repeatedly trained batch versions and the earlier online forest comparison.

The main weakness appears when many features are irrelevant because Mondrian split geometry does not inspect labels.

10×+
paper-reported training-time advantage in several tests
O(N log N)
overall balanced-tree processing cost stated in the paper

Relative training-time index

Lower is better. This is a schematic display of the order reported in the paper.

Accuracy as more training data are seen

A paper-style schematic, not digitised measurements from Figure 3.

22. Two practical views of online performance

These charts reproduce the useful visual ideas from the supplied infographic. Their values are illustrative examples, not measurements reported by the 2014 paper.

Convergence over a data stream

A lower error curve means the model reaches useful performance after fewer incoming observations.

Training time per operation

The comparison separates initial fitting, a small online update, full batch retraining and prediction.

Use these charts as conceptual teaching aids. For dissertation results, replace all values with measured timings from your own code and hardware.

23. The DNA experiment: irrelevant features change the result

This experiment makes the main weakness of label-independent splitting easy to see.

Original DNA input
180
features

Random cuts can spend many splits on dimensions with little predictive value.

feature selection
Filtered experiment: MF†
60
most relevant features

Filtering irrelevant dimensions improves both MF and the closely related ERT-1 model.

Main Figure 3 result

On DNA, plain MF and ERT-1 trail the label-guided batch forests. The paper links this weakness to the presence of irrelevant dimensions.

Appendix F comparison

Dynamic trees outperform plain MF on DNA, but MF† reaches similar test accuracy after the feature filter. MF performs better than dynamic trees on USPS, Satimages and Letter.

Practical lesson

A Mondrian Forest may need careful feature work when the input contains many weak or irrelevant variables. Fast online updates do not remove the need for sensible data representation.

24. What a Mondrian node must store

The tree needs geometric, structural and label information for later updates and predictions.

One node

Bounding box

Lower and upper values for every feature: roughly $2D$ numbers.

Split metadata

Split dimension, split location and split time.

Label statistics

Class counts and the values used by the smoothing approximation.

Tree links

Parent and child references, plus node or leaf status.

A rough engineering view

memory $\propto M \times \text{nodes per tree} \times (2D + K + \text{metadata})$

This expression is an explanatory approximation. Exact memory depends on the code, data types, sparse structures and implementation choices.

$D$

Feature dimension

Bounding-box storage and geometric calculations grow directly with the number of features.

$M$

Number of trees

Every additional tree carries its own structure, boxes and label statistics.

What the original paper states

The authors note that computation is linear in the number of dimensions because rectangles are represented explicitly. They identify high-dimensional data as a limitation for future work.

Do not treat an invented memory curve as evidence. The 2014 paper reports timing and accuracy experiments, but it does not publish a general RAM-versus-stream-size law.

25. Memory growth as the stream expands

The shape below reproduces the requested infographic. The values describe a hypothetical implementation rather than a result from the Mondrian Forest paper.

Illustrative model-state memory

Actual memory depends on the number of trees, feature dimension, lifetime setting, pausing rule, class count and implementation.

More trees

Memory scales approximately with the number of independent trees.

More dimensions

Each node stores lower and upper extents across the feature dimensions.

Deeper partitions

Additional internal nodes store geometry, times and label statistics.

26. Conceptual multi-metric profile

A radar chart compresses several qualitative judgements into one picture. It is useful for orientation, but it is not a scientific benchmark.

Mondrian Forest at a glance

The profile gives high scores to online updating and computational speed. Memory efficiency and ease of implementation depend strongly on tree count, dimension, code quality and comparison method.

Predictive performance: close to batch forests on several datasets in the paper.

Update efficiency: strong when observations arrive sequentially.

Memory: node boxes, split times and class statistics grow with the tree.

Interpretation rule: the polygon is a visual summary of stated judgements. It should never replace measured comparisons.

27. Strengths, limitations and common misunderstandings

Use these specific points when judging whether the method suits a practical streaming task.

Strength

Efficient incremental updates

New observations modify existing trees sequentially. The algorithm traverses short paths from the root instead of discarding and replacing the entire structure.

Strength

Batch-online agreement

The final online tree follows the exact same mathematical probability distribution as a tree generated via batch construction on the identical dataset.

Strength

Probabilistic predictions

The hierarchical label mechanism effectively smooths sparse leaf estimates. This prevents the model from generating erratic probabilities on small samples.

Limitation

Irrelevant features

The random, label-independent splitting mechanism struggles heavily with noise. The algorithm wastes calculations slicing dimensions that contain no actual predictive power.

Limitation

High dimensionality

The bounding boxes must permanently store the specific lower and upper coordinate values for every feature dimension. This mechanism creates a severe computational load in high-dimensional datasets.

Limitation

Concept drift

The foundational algorithm simply aggregates knowledge. Incremental learning alone does not automatically detect drift or decide when outdated historical patterns should be forgotten.

Misunder-
standing

Not standard Random Forest trees

Mondrian trees implement a fundamentally different random partition rule. You cannot achieve a Mondrian Forest by simply feeding streaming data into standard CART algorithms.

Misunder-
standing

Not gradient boosting

These trees function as parallel ensembles. They do not sequentially correct the specific residual errors of previous trees in the way that XGBoost or LightGBM operate.

28. Practical reading checklist

Use these specific questions to verify your understanding while studying the source paper and codebase.

Foundations

  • Can I explain a leaf node as a specific rectangle in feature space?
  • Can I clearly distinguish bagging, standard Random Forests, and boosting algorithms?
  • Can I explain batch learning, online learning, and drift-aware learning as separate operational concepts?

Mondrian geometry

  • What precisely do the node bounding box, split dimension, split location, and split time represent in the model?
  • How does the lambda parameter mathematically affect overall tree complexity?
  • Why are longer dimensions statistically more likely to be split by the exponential clock?

Online extension

  • How does the algorithm calculate the extra extent for an arriving point?
  • Under what exact condition is a new parent inserted above an old node?
  • What specifically is meant by the phrase "same batch and online distribution"?

Prediction and evaluation

  • Why is mathematical smoothing necessary at small leaves?
  • How does the model formulate a prediction for a test point that falls entirely outside the training extent?
  • Which specific error metrics should be tracked continuously through time rather than measured only at the end of the stream?

Suggested first coding exercise for the dissertation

Generate a simple, two-dimensional numerical stream in Python. Plot the geometric partition rules after selected data updates. Following this, record the prequential accuracy and the sequential update time. Compare these recorded metrics against a standard batch Random Forest that is periodically retrained on the identical data stream.

29. One-page summary and references

The minimum mental model to retain before proceeding.

Decision tree
hierarchical axis-aligned regions
Random forest
many varied trees averaged
Mondrian tree
random timed partition that can extend
Mondrian forest
online ensemble of Mondrian trees

Central paper claim

A Mondrian tree can be extended one point at a time so that its mathematical probability distribution after the update is strictly identical to a tree sampled in batch from all points seen so far. This explicit geometric property produces an efficient online forest with competitive accuracy across the authors' standard experiments.

Primary and follow-up sources

  1. Lakshminarayanan, B., Roy, D. M. and Teh, Y. W. (2014). Mondrian Forests: Efficient Online Random Forests. NeurIPS 27, 3140-3148.
  2. Roy, D. M. and Teh, Y. W. (2009). The Mondrian Process. NeurIPS 21.
  3. Mourtada, J., Gaïffas, S. and Scornet, E. (2017). Universal consistency and minimax rates for online Mondrian Forests. NeurIPS 30.
  4. Mourtada, J., Gaïffas, S. and Scornet, E. (2021). AMF: Aggregated Mondrian Forests for Online Learning. Journal of the Royal Statistical Society B.
  5. River documentation and release notes for current Mondrian tree implementations in Python. Consult the package version details carefully before beginning code implementation.

Prepared as explanatory study notes. Diagrams are original simplifications; technical mathematical claims should always be checked directly against the cited source papers.

Appendix: current dissertation snapshot

This page records the position described in the supplied project notes. It does not define the final dissertation.

Already decided

Project area

P2 Statistical Machine Learning.

Supervisor

Dr Yordan Raykov.

Current preference

The Mondrian Forest direction is preferred over the BART direction.

New project plan status

A fresh plan is useful for discussion with the new supervisor, but it is not formally assessed.

Still open
?

Precise research problem

The exact question has not yet been agreed.

?

Type of work

The balance between methodological study, software work and application remains unclear.

?

Data and comparisons

The dataset, comparison models and exact experiments have not yet been selected.

?

Image example

It remains unclear whether background and lighting changes are the central application or only an illustration of streaming data.

Paper

The 2014 article is a starting source.

Area

Online tree-based statistical machine learning.

Final scope

Still requires agreement with the supervisor.

The method has been chosen as the preferred direction. The final dissertation question has not yet been fixed.