Structured Trajectory Serialization

• Structured trajectory serialization is a method that encodes time-ordered spatial positions into compact, queryable formats to support various spatiotemporal queries.
• It employs techniques such as relative compression, differential logs, and grammar-based methods to optimize both storage efficiency and query speed.
• The approach underlies applications in urban mobility analytics, agent-based modeling, and privacy-preserving data analysis, backed by theoretical guarantees on performance.

Structured trajectory serialization refers to the systematic process of encoding trajectories—time-ordered sequences of spatial positions of moving objects—into compact, queryable, and semantically organized representations. This approach enables the efficient storage, transmission, retrieval, and analysis of large-scale spatiotemporal datasets while preserving the spatiotemporal structure necessary for advanced analytic and learning tasks. The defining principle is not just compression, but the imposition of an organization (serialization) on trajectory data that supports a spectrum of spatiotemporal queries and learning objectives.

1. Foundational Models and Principles

Structured trajectory serialization builds on discrete models of object movement. Formally, a trajectory for object $o$ can be written as $$T^{(o)} = \langle(t_1, x_1, y_1), (t_2, x_2, y_2), \dots, (t_n, x_n, y_n)\rangle$$ with time-ordered, typically regularly spaced timestamps and integer-valued coordinates on a spatial grid.

Key principles in structured serialization include:

• Discrete, regularized representation: Trajectories are mapped to grids (spatial or spatiotemporal) and indexed by regular times or fixed aggregation windows.
• Hybrid decomposition: Many approaches separate trajectory serialization into absolute position "snapshots" (periodic images of all objects) and "logs" of movements between snapshots.
• Query-awareness: Structures are optimized to answer point, range, and time-interval queries efficiently according to typical use cases in geographic databases, mobility mining, or agent-based modeling (Brisaboa et al., 2018, Brisaboa et al., 2017, Brisaboa et al., 2016, Tsiligkaridis et al., 2024).

2. Compression and Serialization Techniques

Numerous data structures have been developed for structured serialization, balancing space, query speed, and expressive power:

• Relative Compression of Trajectories (RCT): Implements a Relative Lempel-Ziv (RLZ) parse where each trajectory is encoded as $z$ phrases, each referencing substrings of an artificial global reference trajectory $R$. Storage per trajectory is $O(z \cdot \log N) + o(N)$ bits, and all phrases are serialized with explicit start positions and auxiliary minimum–maximum bounding arrays. RCT supports $O(1)$ access to any object's position at any time and efficient subtrajectory and interval queries (Brisaboa et al., 2018).
• ContaCT: Encodes between-snapshot movement as differential logs, with Elias–Fano encoding for signed differences, and organizes logs with a perfect binary tree where each internal node stores minimum bounding boxes for pruning interval queries. ContaCT's structure achieves $\approx2$ bits/point on real datasets and supports $O(1)$ retrieval at any time point via rank/select over bitmaps (Brisaboa et al., 2017).
• Grammar-based Compression (GraCT): Employs the Re-Pair grammar on movement codewords between snapshots, enriching each nonterminal rule with the temporal span, net displacement, and bounding box. Combined with a $k^2$-tree for snapshots, this grammar-based serialization enables skipping and pruning during queries, with compressed representations at 35–39% the size of raw data and direct in-memory querying (Brisaboa et al., 2016).
• Sequence Transformer for Agent Representation Encodings (STARE): Serializes each agent's trajectory as a sequence of discrete spatial and temporal symbols, tokenized and fed to a Transformer encoder. The output is a sequence of contextualized embeddings—one per location–time token—which is the serialized representation used in supervised or self-supervised downstream tasks (Tsiligkaridis et al., 2024).

3. Auxiliary Structures for Queryability

To ensure serialized trajectories are not only compressed but also queryable in sublinear time, these frameworks attach auxiliary data:

• Rank/Select Bitmaps: Mark phrase or movement boundaries to enable $O(1)$ phrase location and subsegment extraction.
• Minimum/Maximum Bounding Boxes (MBBs) and Range Minimum Query (RMQ) Structures: Support rapid pruning of search intervals in time-window or spatial region queries.
• Spatial Snapshots via $$T^{(o)} = \langle(t_1, x_1, y_1), (t_2, x_2, y_2), \dots, (t_n, x_n, y_n)\rangle$$0-trees: Represent the absolute position of all objects at regular intervals using highly compact, pointerless tree layouts supporting $$T^{(o)} = \langle(t_1, x_1, y_1), (t_2, x_2, y_2), \dots, (t_n, x_n, y_n)\rangle$$1 navigation via rank/select and direct inversion from object IDs to grid cells.
• Unary/Prefix-Sum Bitmaps: Allow for instantaneous computation of cumulative movement or coordinate deltas over arbitrary ranges.

The compressed layout and these indexes enable fast resolution of spatiotemporal queries, with complexity dominated by output or candidate size (Brisaboa et al., 2018, Brisaboa et al., 2017, Brisaboa et al., 2016).

4. Algorithmic Support for Spatiotemporal Queries

Structured trajectory serialization schemes support a suite of standard spatiotemporal queries directly over compressed representations:

• Object-time (point) lookup: $$T^{(o)} = \langle(t_1, x_1, y_1), (t_2, x_2, y_2), \dots, (t_n, x_n, y_n)\rangle$$2 retrieval of an object's coordinates at time $$T^{(o)} = \langle(t_1, x_1, y_1), (t_2, x_2, y_2), \dots, (t_n, x_n, y_n)\rangle$$3 using partial sums or rank/select-enabled movement logs.
• Subtrajectory queries: Extraction of a time-interval subtrajectory in time proportional to its length or less if phrase/skipping or bounding box pruning applies.
• Time-slice (region at single time): Candidate filtering using expanded region search from the last snapshot, followed by $$T^{(o)} = \langle(t_1, x_1, y_1), (t_2, x_2, y_2), \dots, (t_n, x_n, y_n)\rangle$$4 point lookups.
• Time-interval (range in spatial window during interval): Hierarchical or phrase-level pruning using bounding boxes and interval trees; worst-case $$T^{(o)} = \langle(t_1, x_1, y_1), (t_2, x_2, y_2), \dots, (t_n, x_n, y_n)\rangle$$5 per candidate, with $$T^{(o)} = \langle(t_1, x_1, y_1), (t_2, x_2, y_2), \dots, (t_n, x_n, y_n)\rangle$$6 candidates from region expansion (Brisaboa et al., 2018, Brisaboa et al., 2017, Brisaboa et al., 2016).

For learned representations (STARE), the serialized embedding sequence can be used in masked modeling, classification, and clustering tasks, with query complexity depending on model architecture rather than explicit data structures (Tsiligkaridis et al., 2024).

5. Empirical Compression Ratios, Speeds, and Trade-offs

Benchmark studies have reported the following empirical results:

Structure	Bits/pt	Point query (μs)	Time-slice (ms)	Time-interval (ms)
RCT	1.4	0.7	1.2	3.5
ContaCT	1.8	0.8	1.1	4.1
GraCT	2.0	2.5	3.8	9.7

RCT achieves 20–30% improved compression over ContaCT and order-of-magnitude speedups over grammar-based GraCT for interval and time-slice queries (Brisaboa et al., 2018). ContaCT’s compression approaches the entropy of plain differential encoding while maintaining $$T^{(o)} = \langle(t_1, x_1, y_1), (t_2, x_2, y_2), \dots, (t_n, x_n, y_n)\rangle$$7 query access and rapid pruning. GraCT’s grammar approach compresses to $$T^{(o)} = \langle(t_1, x_1, y_1), (t_2, x_2, y_2), \dots, (t_n, x_n, y_n)\rangle$$840% of raw size and supports sub-millisecond object lookups (Brisaboa et al., 2016).

For semantically structured, Transformer-based serialization (STARE), experiments report:

• Agent label classification accuracy exceeding 94% on small synthetic datasets and 88–89% on large-scale simulations, outperforming LSTM and BiLSTM baselines.
• Masked location modeling token recovery rates of 77–85% depending on dataset.
• Embedding clustering reveals semantically meaningful groups of agents or locations (Tsiligkaridis et al., 2024).

A fundamental trade-off in all approaches is between compression ratio, update abilities, and query time. Snapshot interval and phrase size parameters must be tuned to specific dataset statistics and query workload (Brisaboa et al., 2018, Brisaboa et al., 2017, Brisaboa et al., 2016).

6. Applications and Developments

Structured trajectory serialization underpins diverse applications:

• Urban mobility analytics: Enabling fast, memory-efficient analysis of vehicular or ship movement data at city or country scales (Brisaboa et al., 2018, Brisaboa et al., 2017, Brisaboa et al., 2016).
• Agent-based modeling: As in STARE, supporting downstream machine learning, behavioral classification, and pattern-of-life analysis from trajectory embeddings (Tsiligkaridis et al., 2024).
• Database systems and retrieval: Powering large-scale spatiotemporal search engines with worst-case and empirical guarantees on space and time.
• Privacy and anonymization studies: Compact, structured records allow systematic analysis and transformation for privacy-preserving release or aggregation.

Recent developments leverage learned, contextualized representations (Transformers) for richer downstream analysis, supporting classification and similarity-based clustering directly on serialized trajectory embeddings (Tsiligkaridis et al., 2024). Classical approaches remain the state of the art for high-throughput, indexable query workloads on very large datasets (Brisaboa et al., 2018, Brisaboa et al., 2017, Brisaboa et al., 2016).

7. Theoretical Guarantees and Limitations

Structured serialization achieves provable space and time bounds. For RCT: $$T^{(o)} = \langle(t_1, x_1, y_1), (t_2, x_2, y_2), \dots, (t_n, x_n, y_n)\rangle$$9 bits for $z$0 points and $z$1 phrases, $z$2 for point queries, $z$3 for time-slice, and $z$4 for interval queries (Brisaboa et al., 2018). ContaCT provides $z$5 position access and effective pruning via bounding box trees, with space close to differential entropy (Brisaboa et al., 2017).

Limitations include the assumption of regularly sampled time, the need for global grid discretization, and the requirement of pre-specified maximum velocities for candidate expansions. Update support is limited compared to uncompressed representations. Learned (Transformer-based) serialization, while offering expressivity and compatibility with neural models, does not natively guarantee sublinear spatiotemporal query bounds without further architectural augmentation (Tsiligkaridis et al., 2024).

A plausible implication is that the choice of serialization approach should be governed by both data statistics (scale, regularity, movement predictability) and target workload (database querying vs. agent-level modeling), as the trade-off surface is not universal.