Situational aspects for ranking of search results

We've heard and read much about various aspects that are considered for personalised search results, such as, query context, user search history to name a few. But there is one more context in the light of which results can be ranked which is not related to query content and/or user search history. It is known as situational context. Examples of situational context can be search request time, location, browser used for placing request, language, device (system/mobile/tabs etc).

Source : https://www.google.co.in/search?q=situational+context&rlz=1C1CHZL_enIN752IN752&source=lnms&tbm=isch&sa=X&ved=0ahUKEwjB-pK4q4PaAhUvTd8KHX_MBmMQ_AUICigB&biw=1366&bih=637#imgrc=aIOTa6jLvm7sTM:

Advantageous situations

Traditionally the algorithms for ranking developed so far have been using Click through rate as a prime
feature in ranking of the results obtained from the search apart from query similarity and user history.
But this information is not useful in many situations. For instance in case of cold start or when user is
typing his first query in each session. Also if the search request is being made on personal device for
personal documents then this information will not serve the purpose as most of the aforementioned
information is not present.

Heuristics

Some of the heuristics that can leverage the benefits of situational aspects are:

1.   Search queries are short and ambiguous in case of mobile device, so device information
can be used to fill the gaps in query understanding for such cases.

2.   Location can help in providing customized results. An example for the same is, if a user
types “amazon” as query then based on the location it can be either directed to amazon website
or information about amazon forest can be presented. Later information is likely to be more
useful for a user residing in Brazil as compared to someone in US for whom first choice would
make much sense.

3.   Another situational aspect is time. Consider a query “reservation”. It is short and ambiguous
as it can mean train reservation, flight reservation, restaurant reservation etc. But ambiguity can
be resolved upto some extent based on date/time of the query request. Say, for instance if the
query has been requested on a weekend then it can probably mean restaurant reservation or on
a weekday during office hours it can mean booking a conference room or train/flight reservation.

Major Issues

Situational context features can be highly sparse in nature. One such case can be for location. There
can be multiple different locations, but the search request corresponds to only a single location.
Additionally there can be multiple values that corresponds to time such as time of day, day of week,
month of year etc. Using all the possible values of situational context can result in increase of dimensionality
of feature vector space.

Keeping this is mind, deep neural networks can be modelled to work on such sparse data and it has been
observed that the model has performed impressively in such a scenario. Basic workflow includes query
content, situational content and document content given as input to predict a relevance score given to
each document for the given query based on which the documents are ranked.

Conclusion

It has been seen that the situational context can help improving the search results of a query. It can provide more customised result for the user even in the case of cold start. The same can be employed for personal search where click through rate and query document matching score is not present.

Refrences

[1] Zamani, Hamed, et al. "Situational context for ranking in personal search." Proceedings of the 26th International Conference on World Wide Web. International World Wide Web Conferences Steering Committee, 2017.