joeldan192/predictive-model-for-the-voting-system-in-se-network

Semester project Spring 2016 as part of Master's program at EPFL. Supervisor: Dr. Elisa Celis. Lab: LCA-3.

predictive-model-for-the-voting-system-in-se-network's Issues

Plot RSS with test data

In Kranks(2) effects model there's evidence that the model is quite relevant at 0.03. However current R^2 is fit on the training data. Plot RSS or R^2 in test data.

Check why qid-age-ans_position is not unique

• can either make them unique or put them together

Extend examination model for votes with Question features

Whether an Answer is voted on or not might be correlated with:

• Common features with Question. (max mutual information with question?)
• Inherent question quality. (higher quality on questions might induce user to examine answers more thoroughly ...thus vote)

Set AnswerId as categorical variables in Dataframe

• This is kind of tricky for the moment. Unstack method won't behave as expected.
  TODO: once this is achieved, there is the issue of choosing the reference parameter for the categorical variable parameters (standard R convention).

ANOVA for votes data

Pre-processing step to fit the Random effects model.

Add 'accepted' in function that calculates rank

rank: accepted -> score -> random

Get first time series Tableau

Votes are sampled per day and per answer.
Goals:
a)
Votes ~ score + age + nb_comments + nb_ans + position
TODO: nb_ans, position

b) For now the actual models is:
Votes ~ Poisson(e^(quality + position_bias+ age + social_influence(score + TODO in goal a) )))

quality parameter is by post (answer)
position_bias by post and by_observation (changes with time)

Exploratory analysis

Reproduce and asses some results in first week's papers.

Features that entail cognitive bias:

• display position
• simple text statistics (length in characters, nbr of words, capitalization/spelling errors)
• nbr of questions present when voter arrives (cognitive load)

Sugested features (from context):

• reputation
• previous voting history

To try:

• diversity: do voters select answers in an heterogeneous manner?
  (define such sets by clustering over bag of w/model profiles)

Account for dependence within Question and within AnsId

Fixed effects (because of plm implementation).
Multinomial model assumens independence across individuals Q-t pairs. TAnd so does
the 'arellano'( vcovHC) correction. Introduce fixed effects for Qid.
Same for AnsId. As an answer moves through time. and Qi-tk is independent of Qi-tj so we
could account for that dependence.
Warning: the number of parameters would then be very large.
Can be done only with random effects with lme4 package. But there are no covariance estimator
corrections working there

AnsId-Age pairs MUST BE UNIQUE

Namely, in nb Answer_votes after merging with comments inconsistent duplicates are generated for
AnsId-Age pairs

Set a couple different reference points for measuring time for votes on answers

• from creation time of question
• from creation time of answer
  EDIT: Times are, currently, only measured from ans creation (for both votes and comments on answers)

Regularize PageViews model

ReScore has very good performance. Test if this result is robust enough with a few different regularizers.

Define "stable" time intervals for votes

"Stable" time intervals: number of answers doesn't change. (sub-divide with rank doesn't change?)
Such intervals would permit a nested mixed effects model:
time intervals *>(evaluate this part, nested or crossed eff) nb_ans > ans_id > rank

Set parameters for k-ranks

Users might optimize k-ranks. (Consecutive groups of answers from first to last).
This implies one group parameter per answer in voting data.

Normalize by classes

Mean values for ReScore(s) and Pbias(s) are too low.
Might consider another normalization strategy...
By classes perhaps(Ans_count)