2011-06-27

Path finding with Neo4j

In my previous post I talked about graphing databases (Neo4j in particular) and how they can be applied to certain classes of problems where data may have multiple degrees of separation in their relationships.

The thing that makes graphing databases useful is the ability to find relationship paths from one node to another. There are many algorithms for finding paths efficiently, depending on the use case.

Consider a graph that represents a map of a town. Every node in the graph represents an intersection of two or more streets. Each direction of a street is represented as a different relationship: two-way streets will have two relationships between the same two nodes, one pointing from the first node to the second, and the other pointing from the second node to the first. One-way streets will have a single relationship. Each street relationship will also have a distance property. Here is what one such map might look like:
Every street is two-way, except for Market Alley and Park Drive. The numbers in parentheses represent the street distance in hundreds of yards (i. e. 3 = 300 yards.)

Here is the code to create this graph in the database, using the REST client I've been working on (connection initialization is not shown):
$intersections = array(
    "First & Main",
    "Second & Main",
    "Third & Main",
    "First & Oak",
    "Second & Oak",
    "Third & Market",
);

$streets = array(
    // start, end, [direction, distance, name]
    array(0, 1, array('direction'=>'east', 'distance'=>2, 'name'=>'Main St.')),
    array(0, 3, array('direction'=>'south', 'distance'=>2, 'name'=>'First Ave.')),
    array(0, 4, array('direction'=>'south', 'distance'=>3, 'name'=>'Park Dr.')),

    array(1, 0, array('direction'=>'west', 'distance'=>2, 'name'=>'Main St.')),
    array(1, 2, array('direction'=>'east', 'distance'=>2, 'name'=>'Main St.')),
    array(1, 4, array('direction'=>'south', 'distance'=>2, 'name'=>'Second Ave.')),

    array(2, 1, array('direction'=>'west', 'distance'=>2, 'name'=>'Main St.')),
    array(2, 5, array('direction'=>'south', 'distance'=>1, 'name'=>'Third Ave.')),

    array(3, 0, array('direction'=>'north', 'distance'=>2, 'name'=>'First Ave.')),
    array(3, 4, array('direction'=>'east', 'distance'=>2, 'name'=>'Oak St')),

    array(4, 3, array('direction'=>'west', 'distance'=>2, 'name'=>'Oak St.')),
    array(4, 1, array('direction'=>'north', 'distance'=>2, 'name'=>'Second Ave.')),

    array(5, 2, array('direction'=>'north', 'distance'=>1, 'name'=>'Third Ave.')),
    array(5, 4, array('direction'=>'west', 'distance'=>2, 'name'=>'Market Alley')),
);

$nodes = array();
$intersectionIndex = new Index($client, Index::TypeNode, 'intersections');
foreach ($intersections as $intersection) {
    $node = new Node($client);
    $node->setProperty('name', $intersection)->save();
    $intersectionIndex->add($node, 'name', $node->getProperty('name'));
    $nodes[] = $node;
}

foreach ($streets as $street) {
    $start = $nodes[$street[0]];
    $end = $nodes[$street[1]];
    $properties = $street[2];
    $start->relateTo($end, 'CONNECTS')
        ->setProperties($properties)
        ->save();
}
First, we set up a list of the intersections on the map, and also a list of the streets that connect those intersections. Each street will be a relationship that has data about the streets direction, distance and name.

Then, we create each intersection node. Each node is also added to an index. Indexes allow stored nodes to be found quickly. Nodes and relationships can be indexed, and indexing can occur on any node or relationship field. You can even index on fields that don't exist on the node or relationship. We'll use the index later to find the start and end points of our path by name.

Finally, we create a relationship for every street. Relationships can have arbitrary data attached to them, just like nodes.

Now we create way to find and display driving directions from any intersection to any other intersection:
$turns = array(
    'east' => array('north' => 'left','south' => 'right'),
    'west' => array('north' => 'right','south' => 'left'),
    'north' => array('east' => 'right','west' => 'left'),
    'south' => array('east' => 'left','west' => 'right'),
);

$fromNode = $intersectionIndex->findOne("Second & Oak");
$toNode = $intersectionIndex->findOne("Third & Main");

$paths = $fromNode->findPathsTo($toNode, 'CONNECTS', Relationship::DirectionOut)
    ->setMaxDepth(5)
    ->getPaths();

foreach ($paths as $i => $path) {
    $path->setContext(Path::ContextRelationship);
    $prevDirection = null;
    $totalDistance = 0;

    echo "Path " . ($i+1) .":\n";
    foreach ($path as $j => $rel) {
        $direction = $rel->getProperty('direction');
        $distance = $rel->getProperty('distance') * 100;
        $name = $rel->getProperty('name');

        if (!$prevDirection) {
            $action = 'Head';
        } else if ($prevDirection == $direction) {
            $action = 'Continue';
        } else {
            $turn = $turns[$prevDirection][$direction];
            $action = "Turn $turn, and continue";
        }
        $prevDirection = $direction;
        $step = $j+1;
        $totalDistance += $distance;

        echo "\t{$step}: {$action} {$direction} on {$name} for {$distance} yards.\n";
    }
    echo "\tTravel distance: {$totalDistance} yards\n\n";
}
Note that in the `findPathsTo` call, we specify that we want only outgoing relationships. This is how we prevent getting directions that may send us the wrong way down a one-way street. We also set a max depth, since the default search depth is 1. Also, note the `setContext` call which tells the path that we want the `foreach` to loop through the relationships, not the nodes. Running this produces the following directions:
Path 1:
    1: Head north on Second Ave. for 200 yards.
    2: Turn left, and continue west on Main St. for 200 yards.
    Travel distance: 400 yards
We only got back one path because, by default, the path finding algorithm finds the shortest path, meaning the path with the least number of relationships. If we switch the to and from intersections, we can get directions back to our starting point:

Path 1:
    1: Head west on Main St. for 200 yards.
    2: Turn left, and continue south on Second Ave. for 200 yards.
    Travel distance: 400 yards

Path 2:
    1: Head south on Third Ave. for 100 yards.
    2: Turn right, and continue west on Market Alley for 200 yards.
    Travel distance: 300 yards
We got back two paths this time, but one of them is longer than the other. How can this be if we are using the shortest path algorithm? The answer is that shortest path refers only to the number of relationships in the path. We didn't tell our path finder to pay attention to the cost of one path over another. Fortunately, we can do that quite easily by telling the `findPathsTo` call to use a property of each relationship when determining which path is the least expensive:
$paths = $fromNode->findPathsTo($toNode, 'CONNECTS', Relationship::DirectionOut)
    ->setAlgorithm(PathFinder::AlgoDijkstra)
    ->setCostProperty('distance')
    ->setMaxDepth(5)
    ->getPaths();
We're telling the path finder to use Dijkstra's algorithm, an algorithm to find the lowest cost path, which is different from the shortest path. Since our map uses distance, the cost of a path is the total of all the distance properties of the relationships in that path. Running the script now gives us only one path, the path with the least distance:
Path 1:
    1: Head south on Third Ave. for 100 yards.
    2: Turn right, and continue west on Market Alley for 200 yards.
    Travel distance: 300 yards
If there were another path with the same total distance, it would also be displayed. Path cost ignores the number of relationships in the path. If there were a path with 3 relationships totaling 300 yards, and another with 1 relationship totaling 300 yards, both of them would be displayed.

Neo4j has several path finding algorithms built in. Some of them are:
  • shortest path - find paths with the fewest relationships
  • dijkstra - find paths with the lowest cost
  • simple path - find paths with no repeated nodes
  • all - find all paths between two nodes
Neo4j also comes with an expressive traversal API that allows you to create your own path finding algorithms for domain specific needs.

2011-06-15

Neo4j for PHP

Update 2011-09-14: I've put a lot of effort into Neo4jPHP since this post was written. It's pretty full-featured and covers almost 100% of the Neo4j REST interface. Anyone interested in playing with Neo4j from PHP should definitely check it out. I would love some feedback!

Lately, I've been playing around with the graph database Neo4j and its application to certain classes of problems. Graph databases are meant to solve problems in domains where data relationships can be multiple levels deep. For example, in a relational database, it's very easy to answer the question "Give me a list of all actors who have been in a movie with Kevin Bacon":
> desc roles;
+-------------+--------------+------+-----+---------+----------------+
| Field       | Type         | Null | Key | Default | Extra          |
+-------------+--------------+------+-----+---------+----------------+
| id          | int(11)      | NO   | PRI | NULL    | auto_increment |
| actor_name  | varchar(100) | YES  |     | NULL    |                |
| role_name   | varchar(100) | YES  |     | NULL    |                |
| movie_title | varchar(100) | YES  |     | NULL    |                |
+-------------+--------------+------+-----+---------+----------------+

> SELECT actor_name, role_name FROM roles WHERE movie_title IN (SELECT DISTINCT movie_title FROM roles WHERE actor_name='Kevin Bacon')
Excuse my use of sub-queries (re-write it to a JOIN in your head if you wish.)

But suppose you want to get the names of all the actors who have been in a movie with someone who has been in a movie with Kevin Bacon. Suddenly, you have yet another JOIN against the same table. Now add a third degree: someone who has been in a movie with someone who has been in a movie with someone who has been in a movie with Kevin Bacon. As you continue to add degrees, the query becomes increasingly unwieldy, harder to maintain, and less performant.

This is precisely the type of problem graph databases are meant to solve: finding paths between pieces of data that may be one or more relationships removed from each other. They solve it very elegantly by modeling domain objects as graph nodes and edges (relationships in graph db parlance) and then traversing the graph using well-known and efficient algorithms.

The above example can be very easily modeled this way: every actor is a node, every movie is a node, and every role is a relationship going from the actor to the movie they were in:
Now it becomes very easy to find a path from a given actor to Kevin Bacon.

Neo4j is an open source graph database, with both community and enterprise licensing structures. It supports transactions and can handle billions of nodes and relationships in a single instance. It was originally built to be embedded in Java applications, and most of the documentation and examples are evidence of that. Unfortunately, there is no native PHP wrapper for talking to Neo4j.

Luckily, Neo4j also has a built-in REST server and PHP is very good at consuming REST services. There's already a good Neo4j REST PHP library out there, but I decided to write my own to get a better understanding of how the REST interface actually works. You can grab it here and all the code examples below are written using it. The concepts can easily be ported to any Neo4j REST client.

First, we need to initialize a connection to the database. Since this is a REST interface, there is no persistent connection, and in fact, no data communication happens until the first time we need to read or write data:
use Everyman\Neo4j\Client,
    Everyman\Neo4j\Transport,
    Everyman\Neo4j\Node,
    Everyman\Neo4j\Relationship;

$client = new Client(new Transport('localhost', 7474));
Now we need to create a node for each of our actors and movies. This is analogous to the INSERT statements used to enter data into a traditional RDBMS:
$keanu = new Node($client);
$keanu->setProperty('name', 'Keanu Reeves')->save();
$laurence = new Node($client);
$laurence->setProperty('name', 'Laurence Fishburne')->save();
$jennifer = new Node($client);
$jennifer->setProperty('name', 'Jennifer Connelly')->save();
$kevin = new Node($client);
$kevin->setProperty('name', 'Kevin Bacon')->save();

$matrix = new Node($client);
$matrix->setProperty('title', 'The Matrix')->save();
$higherLearning = new Node($client);
$higherLearning->setProperty('title', 'Higher Learning')->save();
$mysticRiver = new Node($client);
$mysticRiver->setProperty('title', 'Mystic River')->save();
Each node has `setProperty` and `getProperty` methods that allow storing arbitrary data on the node. No server communication happens until the `save()` call, which must be called for each node.

Linking an actor to a movie means setting up a relationship between them. In RDBMS terms, the relationship takes the place of a join table or a foreign key column. In the example, the relationship always starts with the actor pointing to the movie, and is tagged as an "acted in" type relationship:
$keanu->relateTo($matrix, 'IN')->save();
$laurence->relateTo($matrix, 'IN')->save();

$laurence->relateTo($higherLearning, 'IN')->save();
$jennifer->relateTo($higherLearning, 'IN')->save();

$laurence->relateTo($mysticRiver, 'IN')->save();
$kevin->relateTo($mysticRiver, 'IN')->save();
The `relateTo` call returns a Relationship object, which is like a node in that it can have arbitrary properties stored on it. Each relationship is also saved to the database.

The direction of the relationship is totally arbitrary; paths can be found regardless of which direction a relationship points. You can use whichever semantics make sense for your problem domain. In the example above, it makes sense that an actor is "in" a movie, but it could just as easily be phrased that a movie "has" an actor. The same two nodes can have multiple relationships to each other, with different directions, types and properties.

The relationships are all set up, and now we are ready to find links between any actor in our system and Kevin Bacon. Note that the maximum length of the path is going to be 12 (6 degrees of separation multiplied by 2 nodes for each degree; an actor node and a movie node.)
$path = $keanu->findPathsTo($kevin)
    ->setMaxDepth(12)
    ->getSinglePath();

foreach ($path as $i => $node) {
    if ($i % 2 == 0) {
        echo $node->getProperty('name');
        if ($i+1 != count($path)) {
            echo " was in\n";
        }
    } else {
        echo "\t" . $node->getProperty('title') . " with\n";
    }
}
A path is an ordered array of nodes. The nodes alternate between being actor and movie nodes.

You can also do the typical "find all the related movies" type queries:
echo $laurence->getProperty('name') . " was in:\n";
$relationships = $laurence->getRelationships('IN');
foreach ($relationships as $relationship) {
    $movie = $relationship->getEndNode();
    echo "\t" . $movie->getProperty('title') . "\n";
}
`getRelationships` returns an array of all relationships that a node has, optionally limiting to only relationships of a given type. It is also possible to specify only incoming or outgoing relationships. We've set up our data so that all 'IN' relationships are from an actor to a movie, so we know that the end node of any 'IN' relationship is a movie.

There is more available in the REST interface, including node and relationship indexing, querying, and traversal (which allows more complicated path finding behaviors.) Transaction/batch operation support over REST is marked "experimental" for now. I'm hoping to add wrappers for more of this functionality soon. I'll also be posting more on the different types of problems that can be solved very elegantly with graphing databases.

The next post dives into a bit more detail about path-finding and some of the different algorithms available.