User behaviour modeling for data prefetching in web applications

A description of method which can be used for reducing the latency in web applications and speculative prefetching. Method proposed by Codete in an internal research project.

Transcript

1. User behaviour modeling for data prefetching in web applications Kacper Łukawski
2. The problem Users tend to use the application in a similar manner over time A majority of modern web application uses multiple HTTP calls to fully generate the content Each request increases total latency perceived by a particular user
3. Mobile applications Network connection is not always good Applications may become useless without an access to some external resources Would be great to have the data locally and synchronize whenever possible
4. User behaviour modeling Get only resources that the user requested for Fetch everything what available Try to predict what might be useful in the nearest future
5. Content based methods Statistical modeling Predicting user's interests
6. Content based methods Statistical modeling Predicting user's interests
7. Content based methods Statistical modeling Predicting user's interests
8. NLP related technique Performs statistical analysis of the language As a base unit of modeling it uses fixed length subsequences Builds statistics of the collocations N-gram statistical model
9. N-gram – an example „The man who does not read has no advantage over the man who cannot read” The man 2 Has no Man who 2 No advantage Who does Advantage over Does not Over the Not read Who cannot Read has Cannot read
10. Statistics are built in the similar manner like it is done in standard n-gram architecture But this time, n is a minimal length of a sequence, not a fixed length Such strategy usually leads to more accurate predictions N-gram+
11. N-gram+ – an example „The man who does not read has no advantage over the man who cannot read” The man who does not read has no advantage over the man who cannot read The man Has no The man who does not read cannot Man who No advantage Man who does not read has no advantage over the man who cannot read Who does Advantage over ... Does not Over the ... Not read Who cannot ... Read has Cannot read
12. A collection of Apache HTTP server logs Delivers the following details of the requests:  Host  Timestamp  Request (URL and method)  HTTP reply code  Number of bytes in the response NASA dataset
13. NASA dataset
14. NASA dataset - results N-gram ~50% accuracy N-gram+ ~60% accuracy
15. Typical web application are built on top of the HTTP protocol HTTP delivers 9 methods – each one has a different meaning (GET, HEAD, PUT, POST, DELETE, OPTIONS, TRACE, CONNECT, PATCH) A proposal of extension
16. GET /order GET /product/{product_id} A proposal of extension – an example GET /order [ {'id': 1, 'products: [{'id': 100}, {'id': 200}]}, {'id': 2, 'products: [{'id': 300}]} ] GET /product/100 { 'id': 100, 'name': 'lorem' } GET /product/200 { 'id': 200, 'name': 'ipsum' } GET /product/300 { 'id': 300, 'name': 'dolor' }
17. Typical API endpoint: GET /resource/{id}/?foo={foo}&bar={bar} Between two HTTP actions from one session: ➔ No relation at all ➔ Request or response tokens of second action are taken from the first one ➔ Tokens are filled using some external knowledge ➔ Value of the token in the second action might be calculated from the values of the first one A proposal of extension - relations
18. Typical API endpoint: GET /resource/{id}/?foo={foo}&bar={bar} Between two HTTP actions from one session: ➔ No relation at all ➔ Request or response tokens of second action are taken from the first one ➔ Tokens are filled using some external knowledge ➔ Value of the token in the second action might be calculated from the values of the first one A proposal of extension - relations
19. Let's treat each HTTP request and response as a single action performed by the user Try to find the relations between the actions that often take place in the similar order Use only the actions that do not change anything (GET, HEAD) A proposal of extension - contd
20. Assign the HTTP actions into the HTTP endpoints that were used to process them Tokenize each request and response Using some n-gram-like architecture, try to find the relations within the subsequences (request/response request) → Collect the statistics of token values in the requests A proposal of extension – an algorithm
21. In the prediction phase try to find the most probable endpoint that will be used in the next step Having the endpoint, fill the tokens using relations to previous actions If not all the tokens can be filled, use the statistics of the values Perform the action(s) using predicted values of tokens Send aggregated responses at once A proposal of extension – an algorithm
22. A. Georgakis, H. Li, “User behavior modeling and content based speculative web page prefetching”, Data & Knowledge Engineering 59 (2006) 770–788 M. Narvekar, S. S. Banu, “Predicting User’s Web Navigation Behavior Using Hybrid Approach”, Procedia Computer Science 45 ( 2015 ) 3 – 12 http://ita.ee.lbl.gov/html/contrib/NASA-HTTP.html https://en.wikipedia.org/wiki/N-gram References