Knowledge Enabled Location Prediction of Twitter Users

Editor's Notes

• #3: “City of Lights” – nickname of Paris. When we see a text like this, we use that background knowledge we have to interpret the text. Similarly, background knowledge can be used to improve a machine’s ability to understand and interpret text.
• #4: Background knowledge is central to the idea of semantic web. Domain specific knowledgebases such as MusicBrainz, UMLS, Dbpedia etc have been used to solve problems such as named entity recognition and disambiguation, relevant document retrieval, data analysis etc. Today I am going to talk about applying background knowledge to predict the geographic location of a user. In this context, the location of a user is their home location at the city level.
• #5: Social media has grown rapidly. Many applications have been developed based on Twitter Associating geographic information with Twitter users can provide value addition to many applications
• #6: Location provides “context”
• #7: Users can volunteer geographic information through cellphone or profile
• #8: Cheng et al. found 21% users contained location as granular as city, state in their profile There was a need to automatically infer the location of a user
• #9: Network based approach uses training data to determine hidden patterns in the communications between a user and his friends. These observations are used to predict the location of a user
• #10: Geographic location of a user influences the contents of their tweets The general idea behind these approaches is to determine the probabilistic distribution of words across a region. Different approaches such as language models, topic models, Gaussian mixture model and ensemble based classifiers have been used for the task of location prediction.
• #11: We address these weaknesses by using a knowledge-based approach to extract location specific concepts from Wikipedia. Content based prediction of the location of a Twitter user by exploiting Wikipedia to create a location specific knowledge-base
• #12: This approach consists of three modules – A knowledge base generator that extracts location specific concepts from a crowd sourced knowledgebase. A User profile generator that creates a semantic profile of a user whose location is to be determined and a Location Prediction module that uses the semantic profile and the knowledgebase to predict the top-k locations of a user.
• #13: Knowledgebase generator extracts location specific concepts from Wikipedia.
• #14: Local Entities: Entities that have a high relatedness to a city and can discriminate between geographic locations
• #15: Wikipedia is publicly available. Anyone can edit a Wikipedia article, correct errors and compensate for any biased views. Wikipedia has been used as background knowledge to solve many problems. At Kno.e.sis, Wikipedia has been used in many applications. Doozer used Wikipedia to create domain specific ontologies, Blooms used Wikipedia for Ontology Alignment and the hierarchical interest graph for a personalization system, was also based on the category structure of Wikipedia. Rephrase this slide.
• #16: So we hypothesize that for each location, the internal links in its Wikipedia page represent entities that are relevant to it and we consider them as Local Entities of the city. In other words, the local entities of a city are all the outgoing links from its Wikipedia page.
• #17: For example, consider this snippet from the Wikipedia page of San Francisco. The San Francisco Chronicle is a major daily newspaper in San Francisco. Clearly it is more local to San Francisco than CNN or MSNBC which is are national media outlets.
• #18: So now our goal is to score each entity with respect to the city such that the score reflects the localness measure of the entity with respect to the city. We experiment with measures of three different classes to score the localness of an entity with respect to a city. That is, association based measure, graph based measure and semantic overlap based measure.
• #19: Association based measure: Compute relatedness based on their occurrences in a large corpus PMI is a measure of how much the actual probability of their occurrence differs from what is expected based on the probabilities of their individual occurrences. Wrt San Francisco: CNN: 7.94 San Francisco Chronicle: 10.41
• #20: Construct a directed graph of local entities for each city Ulrik Brandes
• #21: BC used to compute the importance of an actor in a social network. The importance is measured in terms of shortest paths. For each node, the BC is the sum of the fraction of shortest paths that pass through that node. So BC is helping us determine which nodes are important in the network of local entities of a city.
• #22: Based on the idea that higher is the overlap between concepts found in the Wikipedia pages of a city and an entity higher is the degree of localness of the entity
• #23: Jaccard Index is a symmetric measure to compute the semantic overlap between a city and an entity. But we find that a local entity generally represents a part of the city. For example, Golden Gate Bridge will not completely overlap with all the concepts of San Francisco which contains entities from different categories like Climate, History, Geography etc. With that in mind, we use Tversky Index which is an asymmetric similarity measure.
• #24: Tversky Index is a unidirectional measure of similarity of the local entity with respect to the city. We penalize the local entity for every concept in its Wikipedia page that is not present in the city
• #25: The next module is the user profile generator. We create a semantic profile of each user consisting of Wikipedia entities found in their tweets
• #26: We create a semantic profile of each user consisting of Wikipedia entities found in their tweets. First step in the creation of User Profile is Entity Linking: Identification of entities from tweets and linking them to their Wikipedia article. Zemanta for Entity Linking
• #27: The next step is Entity Scoring i.e. Scoring each local entity in a user’s tweet using frequency of their occurrence. We are predicting the home location of a user based on a set o their historic tweets. Using the frequency helps to determine how relevant the entity is with respect to the user. For example: a football fan may tweet about many football teams, chances are that he will tweet more frequently about the football team of his city
• #28: Brief description of the approach
• #29: For predicting the location of a user we compute an aggregate score for each city whose local entities are found in a user’s tweets.
• #30: We first created a knowledgebase of locations that contains location specific entities along with a score that represents the localness of the entity with respect to the city. Next, for a user whose location is to be predicted we create a profile that consists of entities mentioned in their tweets and their frequency. Finally, we predict the top-k locations of a user.
• #31: We evaluated our approach on a test dataset published by Cheng et al. Their test dataset consists of 5119 active users from continental United States with approximately 1000 tweets per user. These users are spread across 569 cities in US. Spammers and bots are cleaned from the dataset to ensure a clean dataset.
• #33: We compute the distance as a straight-line distance between a pair of lat-longs using haversine formula
• #35: PMI – not normalized, sensitive to count of occurrences
• #36: We find that betweenness centrality does a good job of assigning low scores to common entities such as the National Weather Service but it fails for entities which have some relevance to the city but no distinguishing factor.
• #37: Endicott, New York is the birthplace of IBM. So the Wikipedia page of Endicott has an entire section dedicated to IBM and this section contains entities very specific to IBM e.g. punched cards and T.J Watson. Because the shortest path between these IBM-specific nodes and other nodes of the city lies through IBM, the betweenness centrality of IBM is very high. In such cases, local entities scored using betweenness centrality lead to incorrect predictions.
• #38: Jaccard Index underperforms for entities with fewer entities than the city.
• #39: Example: Eureka Valley is a residential neighbourhood of San Francisco so we would expect it to be more local to SFO than California.
• #40: With Tversky, the localness of a local entity only diminishes for entities in its page not present in the city.
• #47: Introduction to Twitter and location of a Twitter user
• #48: We find that the more local entities we find in a user’s tweet the higher is our ability to predict their location accurately. This is quite intuitive. If you tweet more and give more clues about your location then there is a higher possibility of being able to locate you. On the other hand, if your tweets are specific to a certain topic such as technology or national level or world politics then it is difficult to find evidence that can be used to locate you accurately. As shown in this graph, the predictions made on the basis of 10 or more entities were able to locate 66% of the users within 100 miles.
• #49: From this dataset we selected users from the top-100 most populated cities of US and found that our algorithm was able to locate 60% of the users within 100 miles of their actual location and 54% users at exactly the city level.