140 likes | 265 Vues
This project presents an interactive recommendation system that suggests travel locations and activities based on user queries. Utilizing an integrated random walk approach, the system constructs a graph where nodes represent locations and activities, and edges signify their relationships. The relevance scores for recommendations are computed similarly to PageRank, ensuring suggestions are both relevant and popular. Key components include natural language processing for query term extraction, TFIDF weighting for similarity computation, and an interactive user interface to enhance the travel planning experience.
E N D
Interactive Recommendation about Location and Activity Using Integrated Random Walk Group Members R98922004 Yun-Nung Chen R98922009 Che-An Lu R99922107 Ring Li R99922109 Jerry Cheng
Goal of Project • For a query about location • Recommend locations/activities • Show the results in the interactive system
Dataset • Text documents about traveling • Crow from PTT-travel
Our Idea • Construct a graph • node: location/activity • edge: relation between two nodes • Compute relevance score • Similar to PageRank
Flow Chart Archive of documents from PTT-travel Similarity Computation user Interactive Recommendation System Location/Activity Extraction Query-related score Computation query
Key Term Extraction • Word Segmentation • PoS Tagging • VERB + NOUN activity • NOUN + web filter location
Flow Chart Archive of documents from PTT-travel Similarity Computation user Interactive Recommendation System Location/Activity Extraction Query-related score Computation query
Similarity Computation • TFIDF Weight The similarity between two nodes (location/activity)
Flow Chart Archive of documents from PTT-travel Similarity Computation user Interactive Recommendation System Location/Activity Extraction Query-related score Computation query
Proposed Approach • Integrated Random Walk The query-related score
Proposed Approach • Integrated Random Walk The propagation about similarity of node with the same type The propagation about similarity of node with different types
Proposed Approach • Integrated Random Walk Similar to PageRank Final scores can represent relativeness to query and popularity
Query Expansion • For unseen query about location • Top 10 snippets TFIDF vector • View the most similar one as the query
Experiments • Set the initial query-related scores • 0.35 for query • 0.15 for location • 0.5 for activity • Set the ratio for random walk • alpha, beta = 0.85 / 2 See the Demo!!!