Modelling Time-aware Search Tasks for Search Personalisation
Thanh VuComputing and Communications DepartmentThe Open UniversityCRC PhD Student Conference, July 2015
Modelling Time-aware Search Tasks for Search Personalisation
Search Personalisation2Return search results based onThe input queryThe user searching interests Different users submit the same input query will probably get different search result lists Even an individual user will get different search results at different search times (e.g., Open US)
Search TaskA search task represents an atomic user information needA user may submit several queries within a search task and handle several tasks within a search session
Mining and modelling search tasks helps to improve the performance of search personalisation3
Research Problem4Previous studies have limitation in modelling the dynamic nature of search tasks With the time change, the search intent and the user interests may also change
Research Questions5How can we model search tasks with time-awareness?Can the time-aware search task help to improve search performance?
Modelling Time-aware Search Tasks (1/2)6Identify search tasks in a search session using the Query Task Clustering approach
Modelling Time-aware Search Tasks (2/2)7Extract latent topics from the tasks clicked documents using Latent Dirichlet AllocationUse a decay function to model a time-aware task
Application to Re-ranking8Utilise these time-aware search tasks to re-rank the original list of documents returned by a commercial search engineThe good search performance means the more relevant document should be returned in the higher rank
Experiment9DatasetThe query logs of 106 anonymous users from a commercial search engine for 15 days from 01 to 15 July 2012The training set contains the log data in the first 9 days and the test set contains the log data in the remaining days
Experiment10Baseline and Personalisation StrategiesDefault: The original ranked results from the search engineLongTerm: Use temporal long-term user profiles constructed using the users whole search historyShortTerm: Use temporal short-term profiles constructed using the users current search sessionStaticTask: Use non-temporal search tasksTimeTask: Use time-aware search tasks
Overall Performance11Evaluation metrics Mean Average Precision (MAP)Precision (P@k)Mean Reciprocal Rank (MRR)Normalized Discounted Cumulative Gain (nDCG@k)For each metric, the higher value indicates the better ranking
Overall Performance12
Using time-aware search tasks, TimeTask achieves significantly better performance than other methodsWith time-awareness, ShortTerm gains advantage over StaticTask (without time-awareness)
12
TakeawaysConstructing the time-aware search task using the topic extracted from the tasks relevant (clicked) documents and a decay functionTime-aware search tasks help to improve the performance of search personalisation
13
Thank You!Any Questions?14
Modelling Time-aware Search Tasks (2/3)15Extract latent topics from the tasks clicked documents using Latent Dirichlet Allocation
Application to Re-ranking (2/2)16Compute Jensen-Shannon divergence between the time-aware search tasks and a returned document dExtract other non-personalised features of the input query q and document d, e.g., the rank of d on the original listApply the learning to rank algorithm, LambdaMART, to train re-ranking models
