Markov Geographic Model (MGM)

• Markov Geographic Model (MGM) is a grid-based, non-parametric online classifier that maps streaming time series into discretized embedded states using derivative delay embedding.
• It combines state geographic distribution with first-order transition probabilities to handle phase shifts, variable lengths, and baseline variations in real time.
• Empirical evaluations on handwriting, action recognition, and sensor data show that MGM achieves high accuracy and constant memory usage, outperforming traditional methods.

Searching arXiv for the specified papers to ground the article and confirm bibliographic details. Markov Geographic Model (MGM) is a non-parametric, grid-based model for online modeling and classification of streaming time series within the DDE-MGM framework. It combines two sources of information derived from derivative delay embedding (DDE): the state geographic distribution, which describes where embedded states occur in a discretized embedding space, and the state transitions, which describe how those states evolve over time. In the formulation introduced in "Derivative Delay Embedding: Online Modeling of Streaming Time Series," MGM is designed to operate incrementally, with approximately constant memory determined by the discretization grid, while remaining phase- and length-invariant for recurring patterns in streaming data (Zhang et al., 2016).

1. Conceptual basis

The defining principle of MGM is the joint use of geometric occupancy and Markovian dynamics. The geometric component models the distribution of states in the discretized embedding space, while the Markov component models transition probabilities between successive states. This differs from traditional Markov models such as Hidden Markov Models, which only capture transition dynamics of the form $P(x_t \mid x_{t-1})$. MGM therefore represents both where a trajectory resides in embedding space and how it moves through that space (Zhang et al., 2016).

Within the DDE-MGM framework, this dual representation is intended to address common difficulties in streaming time-series analysis, including variable length, misalignment, and baseline shifts. The paper states that ordinary delay embedding can produce similar-shaped trajectories that are differently positioned, whereas derivative delay embedding maps such cases to identical trajectories in embedding space. This suggests that the term geographic refers not to physical geography, but to the spatial distribution of embedded states in the discretized state space.

The model is explicitly described as non-parametric and online. It does not assume a parametric structure, and its statistics are updated incrementally as new samples arrive. The paper further characterizes it as less sensitive to initialization and to individual starting points or outliers, because it leverages the overall trajectory rather than relying primarily on an initial-state distribution (Zhang et al., 2016).

2. Derivative delay embedding and state construction

MGM is defined on top of derivative delay embedding. For a multivariate time series, DDE produces an embedded sequence

$$\Phi'(x_t) = G(y'_t, y'_{t+s}, \ldots, y'_{t+(d-1)s}),$$

where $y'_t = y_t - y_{t-1}$, and $G(\cdot)$ discretizes the continuous embedding space into a finite grid (Zhang et al., 2016).

This construction serves two purposes. First, the derivative operation is used to remove baseline effects and improve invariance to phase and alignment. Second, discretization converts the embedding into a finite set of grid cells that can be counted and updated online. Each occupied grid cell becomes a state for MGM, and a streaming trajectory is represented as a sequence of visits to these states.

The paper emphasizes that DDE-MGM makes no assumptions about fixed length or alignment, and therefore does not require preprocessing, segmentation, or normalization of the raw streaming data. After the initial sample is processed, modeling and classification can be performed in real time and in parallel. A plausible implication is that the embedding stage is not merely a preprocessing convenience; it is the mechanism that renders the subsequent state model suitable for arbitrary and potentially infinite streams (Zhang et al., 2016).

3. Probabilistic formulation

MGM consists of two probability models. The first is the state geographic distribution, which assigns a probability to each discretized embedded state according to its visit count:

$$P(x_t) = \frac{\log\left(|\Phi'(x_t)| + 1\right)}{\sum_i \log\left(|\Phi'(x_i)| + 1\right)}.$$

Here, $|\Phi'(x_t)|$ denotes the number of visits to state $x_t$. The logarithm is used to suppress the influence of states that are visited very frequently because of baseline or noise effects (Zhang et al., 2016).

The second is the state transition model, which estimates a first-order transition probability between successive embedded states:

$$P(x_t \mid x_{t-1}) = \frac{|\Phi'(x_t); \Phi'(x_{t-1})|}{\sum_i |\Phi'(x_i); \Phi'(x_{t-1})|},$$

where $|\Phi'(x_t); \Phi'(x_{t-1})|$ is the count of observed transitions from $x_{t-1}$ to $$\Phi'(x_t) = G(y'_t, y'_{t+s}, \ldots, y'_{t+(d-1)s}),$$0 (Zhang et al., 2016).

For classification, a test trajectory $$\Phi'(x_t) = G(y'_t, y'_{t+s}, \ldots, y'_{t+(d-1)s}),$$1 is scored against each class-specific MGM by combining state-distribution and transition contributions. The paper separates these as

$$\Phi'(x_t) = G(y'_t, y'_{t+s}, \ldots, y'_{t+(d-1)s}),$$2

and

$$\Phi'(x_t) = G(y'_t, y'_{t+s}, \ldots, y'_{t+(d-1)s}),$$3

with overall similarity formed from both terms. Classification is then performed by assigning the most likely MGM, that is, the class whose model yields the highest similarity score (Zhang et al., 2016).

4. Incremental operation and neighborhood matching

In training, MGM is updated point by point as the stream arrives. For each new embedded point, the model increments the count of the corresponding grid cell for the state distribution and increments the count of the transition from the previous grid cell to the current one for the Markov model. This enables online modeling with a fixed, small memory buffer rather than a retained history of the full stream (Zhang et al., 2016).

In classification, each incoming test window is transformed by DDE and discretized. For each class MGM, the model retrieves the relevant state and transition probabilities and computes the combined similarity, updating incrementally as new points arrive. The paper describes this as true online modeling and classification, including for data streams of arbitrary and infinite length.

A further refinement is neighborhood matching. Exact cell matching can be brittle when a test trajectory falls near but not exactly on cells observed during training. MGM therefore pools contribution from a small-radius neighborhood around the current state, and for transition similarity it sums counts over neighborhoods of both the current and previous states. The stated purpose is improved robustness to noise (Zhang et al., 2016).

The implementation is designed to preserve online efficiency. The state distribution map is typically a 2D or $$\Phi'(x_t) = G(y'_t, y'_{t+s}, \ldots, y'_{t+(d-1)s}),$$4-dimensional array, while transitions are stored in an expandable sparse list rather than in a full, mostly empty matrix. The memory footprint is therefore approximately constant and determined by the grid size of discretization, not by stream length (Zhang et al., 2016).

5. Invariances, robustness, and distinctions from related Markov approaches

MGM is presented as robust to phase, length, and baseline variation because recurring patterns map to the same DDE-MGM signature. The paper’s illustrative example states that two time series with the same pattern but different baseline or phase produce similar-shaped but differently positioned trajectories under ordinary delay embedding, but identical trajectories after DDE. On that basis, MGM is described as immune to baseline and alignment changes when trained on the discretized DDE representation (Zhang et al., 2016).

The model also differs from classical Markov approaches in several explicit ways: it includes geographic information in addition to transition probabilities; it is non-parametric; it is designed for incremental online updates; it uses grid discretization to impose constant memory need; and it incorporates neighborhood matching for adaptability to small variations and noise. The paper further states that MGM is less sensitive to initial-state probability than traditional models (Zhang et al., 2016).

A recurrent source of terminological confusion is the word geographic. In MGM, the state space is the discretized embedding space produced by DDE, not a physical road network or geographic coordinate system. This becomes clearer when MGM is contrasted with next-location prediction models such as NLPMM, in which each state corresponds to a sampling location, transitions correspond to observed movements from one location to another, and global and personal Markov models are integrated using linear regression for trajectory prediction (Chen et al., 2020). MGM, by contrast, is a streaming time-series classification model over embedded states rather than a predictor of future physical locations.

6. Empirical performance and application domains

The experimental evidence reported for DDE-MGM covers handwriting, action recognition, and wearable sensor data. The paper states that DDE-MGM outperforms both offline methods, including HMM and DTW, and online methods such as kernel-based perceptrons and online gradient descent, particularly on unaligned and variable-length datasets. It also reports modeling and classification rates exceeding 11,000 Hz with constant memory, making the method suitable for real-time or embedded deployment (Zhang et al., 2016).

Dataset	DDE-MGM result	Comparative note
UCI Character Trajectory	95.45% in online mode	Highest or comparable accuracy, competitive runtime
MSR Action3D	93.04% offline; 79.37% online	Other online methods achieve only $$\Phi'(x_t) = G(y'_t, y'_{t+s}, \ldots, y'_{t+(d-1)s}),$$5
PAMAP Outdoor Activities	86.96%	Best accuracy; memory consumption $$\Phi'(x_t) = G(y'_t, y'_{t+s}, \ldots, y'_{t+(d-1)s}),$$6 for streams $$\Phi'(x_t) = G(y'_t, y'_{t+s}, \ldots, y'_{t+(d-1)s}),$$7

On the UCI Character Trajectory dataset, which is described as well aligned and of similar length, DDE-MGM achieves 95.45% in online mode. On MSR Action3D, which is variable length and unaligned, it achieves 93.04% offline and 79.37% online, while the paper notes that other online methods achieve only approximately 30% because of their reliance on alignment or truncation. On the PAMAP Outdoor Activities dataset, which is described as long, repetitive, and unsegmented streaming data, DDE-MGM achieves 86.96% accuracy with memory consumption below 7KB for streams larger than 10MB (Zhang et al., 2016).

These results support the paper’s central characterization of MGM as an online classifier for streaming time series that couples occupancy statistics and transition dynamics in a discretized derivative delay embedding. A plausible implication is that the method’s practical value arises less from any single probabilistic component than from the interaction among DDE, discretization, sparse online counting, and neighborhood-based robustness.