Traffic forecasts can get sharper when sensors learn what nearby places mean

How public place knowledge becomes a meaning layer for each sensor

A traffic sensor sits in a particular corner of a city, and that location carries meaning beyond the road it watches. Wikidata, a large public knowledge base that stores facts about places, institutions, landmarks, and administrative areas, supplies the raw material. This choice matters because building a separate place dictionary for every city would be slow, costly, and hard to reuse elsewhere.

For each sensor, the first step is to collect nearby entries from Wikidata inside a chosen radius. A 600 meter search captures immediate street level surroundings, such as intersections, road infrastructure, and small local features. A 2.5 kilometer search reaches farther out and includes major attractors and district scale effects that can shape traffic over a wider area. The selected places form a local semantic neighborhood for the sensor.

Two versions of this neighborhood are used. The direct version keeps only the nearby entries and their type information. The one hop version adds connected entities, so a place can bring in related facts from one step away. This makes the context broader while still staying tied to the sensor’s surroundings. At 600 meters, the direct subgraph contained 797 entries and 1827 links; adding one hop expanded it to 2920 entries and 12523 links. At 2.5 kilometers, the direct version had 2770 entries and 8062 links, while the one hop version reached 29655 entries and 144141 links.

Those local webs are then turned into numbers. A standard knowledge graph embedding method represents each entry as a complex valued vector, which can be thought of as a point with two coordinated values rather than a single number. This form is useful because many relationships in public place data are directional or asymmetric. Each nearby place therefore contributes its own numerical signature to the sensor’s profile.

The sensor receives one semantic vector by averaging all associated place vectors. Averaging produces a fixed size representation that does not depend on how many places were found or on their order. It also smooths mixed surroundings, so a business district with schools, shops, transit stops, and offices becomes one blended signature rather than a list of separate labels. All embeddings are set to 64 dimensions after preliminary tests, balancing expressive power against computational cost. They are trained self supervised on the extracted place facts, aiming for coherent representations of places and entities instead of tuning directly toward traffic prediction.

Turning that meaning into connections between sensors

Once each sensor has a compact description of its surroundings, the method compares those descriptions pairwise. It turns meaning into links by asking whether two sensors look alike from the point of view of their nearby places. The comparison uses a cosine-like score, adapted for vectors that contain complex numbers. Each sensor’s semantic vector is treated like an arrow, and the score measures how much two arrows point in the same direction. Because complex vectors carry more than ordinary length, the method takes the real part of their inner product and then divides by both vector lengths. This gives a similarity value that is high when two places share functional character, even if they are far apart. The resulting table has one row and one column for every sensor, with a weight in each cell. That table is the new adjacency matrix, or map of connections. It is not a purely abstract place map. Nearby sensors often share the same streets, shops, and landmarks, so their semantic profiles tend to be similar as well. This spatial overlap keeps the new links grounded in geography. At the same time, two distant business districts can receive similar weights because their surroundings play similar roles. A resort area in one part of a city may connect to another resort area, not simply to the nearest highway sensor. The matrix therefore mixes local geography with functional resemblance. Each connection is a soft clue, not a hard rule about which sensors influence each other. Forecasting models can use this map as an extra channel for spreading information across the sensor network. By doing so, they gain a way to notice that traffic behavior may depend on place type, not only road distance. The whole step is deliberately simple: compare vectors, score pairs, and fill a matrix. Yet that simplicity matters because it lets existing traffic models receive richer context without being rebuilt from scratch.

Replacing the road map with the place-based map

When the road map is set aside and replaced with a place based map, the result is mixed. In this test, traffic forecasters receive connections drawn from nearby places instead of distances along streets. Some models like that change, while others become less accurate.

The clearest gain came from D2STGNN, which cut mean absolute error by 15.2 percent. That improvement used the 600 meter direct neighborhood matrix, so it linked sensors through their immediate surroundings. STGCN improved by up to 6.2 percent, and STTN by up to 2.7 percent.

Those gains are real, but they do not prove that place meaning caused them. A random connection pattern with the same density also helped several models. D2STGNN gained 13.2 percent from that random map, and STGCN gained 3.7 percent. This means part of the benefit may come from a different graph shape, not from meaningful place links.

Two models resisted replacement more strongly. Fixed GWaveNet worsened with every semantic matrix, by as much as 31.4 percent in mean absolute error. Adaptive GWaveNet also slipped, though only about 1.7 to 2.1 percent. DCRNN behaved less predictably, improving with one random setup and one semantic setup but worsening by up to 24.9 percent elsewhere.

The replacement test therefore shows a useful pattern: some forecasters can use place based links as their main spatial map. Others need the road network to stay stable, and forcing them onto a semantic graph alone makes predictions worse. That result sets up a harder question: whether place meaning can help without erasing the street structure that traffic already follows.

Keeping the road map and adding the place-based map

The added-map test kept the ordinary road-distance connections and gave each forecaster a second connection map at the same time. A second map linked sensors by the character of their surroundings, not just by how far apart they sat on the highway network. In this setup, the model could use both views of space while learning how much weight to give each one. The design helps because traffic moves along roads and also responds to what a place is for, from office blocks to shopping strips.

The San Diego test used 716 sensors and more than 35,000 time measurements, so the result was not drawn from a tiny corner of one city. Most forecasters improved when the place-based map was added beside the road-distance map. STGCN cut its mean absolute error by 7.7 percent with the wider 600-meter neighborhood version. D2STGNN did even better, lowering that same error measure by 12.4 percent under the same setting. STTN gained 5.0 percent when it used the broader 2.5-kilometer direct-neighborhood map.

DCRNN showed the clearest warning, because its response depended heavily on which place-based map it received. With the wider one-hop maps, it cut error by 15.9 percent at 600 meters and by 11.7 percent at 2.5 kilometers. The direct-neighborhood maps produced the opposite result, raising error by 114.8 percent at 600 meters and by 77.5 percent at 2.5 kilometers. That swing shows the added map is not a free upgrade for every architecture.

A map that understands where it is

The semantic layer gives traffic models a second way to see the city. Instead of treating sensors as points joined by road distance, it treats them as places with meaning. A sensor near offices, schools, or transit hubs carries a different local signature from one beside a residential street. That signature comes from Wikidata, so the same procedure can move across cities without hand building a local map.

The forecast gains were real, and they were uneven across models. Some forecasters improved when the semantic matrix was added to the road network, while others changed little or worsened. A random connection pattern also helped several models, which means part of the benefit may come from a different graph shape rather than from meaning alone.

The main opportunity is that the method can move across cities without rebuilding a local place ontology by hand. It also opens a path for forecasters to use administrative roles and nearby functions without rebuilding their core architecture. The practical obstacles are also clear enough to shape future tests. Wikidata is broad, but it may miss local details or encode places unevenly. The semantic matrix also carries spatial signal, so a gain cannot always be traced to meaning alone. Some models became much worse with certain configurations, which warns that adding context can disturb an architecture instead of helping it.

Beyond traffic prediction, the idea points to a larger shift in how machines read cities. A street carries more than geometry; it also has a role in daily life. If forecasters can learn that role from shared place knowledge, they may anticipate congestion before road distance alone explains it. Can a city become predictable through the meaning of its places rather than the geometry of its streets?