Date post: | 12-Jul-2015 |
Category: |
Technology |
Upload: | moresmile |
View: | 69 times |
Download: | 4 times |
ABSTRACT
Tenure
what to recommend to a user when to make appropriate recommendations and its
impact on the item selection in the context of a job recommender system
The proportional hazards model, hierarchical Bayesian framework
estimates the likelihood of a user’s decision to make a job transition at a certain
time, which is denoted as the tenure-based decision probability
CONTRIBUTION
Analyze the problem of finding the right time to make recommendations in the job
domain.
Propose using the proportional hazards model to tackle the problem and extend it with
a hierarchical Bayesian framework.
Evaluate the model with a real-world job application data from LinkedIn
RELATED WORK
Major recommendation approaches: content-based filtering and collaborative
filtering
Timeliness
E.g. software engineer – senior software engineer
Method
HIERARCHICAL PROPORTIONAL HAZARDS MODEL
Problem Definition
Review of Proportional Hazards Model
Model Extension with Bayesian Framework
Parameter Estimation
TENURE-BASED DECISION PROBABILITY
HIERARCHICAL PROPORTIONAL
HAZARDS MODEL
Goal
Predict the probability that a user makes a decision of item jb at current time tb ,given
that she made the last decision of ja at time ta and she did not make the transition
decision up to time tb .
HIERARCHICAL PROPORTIONAL
HAZARDS MODEL
Review of Proportional Hazards Model
Survival function :determines the failure of an event
Failure: as a user making a decision to transit to a new job
HIERARCHICAL PROPORTIONAL
HAZARDS MODEL
Two common approaches to incorporate covariates x in the hazards model:
Cox proportional hazards model :
Accelerated life model :
EVALUATION OF HAZARDS MODEL
Models to Compare
H-One :the hazards model that fits a single set of parameters with no
covariates ;m = {∗ → ∗}
H-Source :the hazards model that fits multiple sets of parameters with no covariates to the tenure data ;m = { a → ∗}
H-SourceDest :m = { a → b }
H-SourceDestCov : incorporates covariates into the hazards model in H-SourceDest .
EVALUATION OF HAZARDS MODEL
covariates
1) about the user u : the user’s gender, age, number of connections, number ofjobs that the user has changed, average months that the user changes a job;
2)about the item ja or jb : discretized company size, the company age
3)about the relationship between ja and jb : the ratio of the company size, the ratio of the company age; whether j a and j b are in the same function, whether they are in the same industry;
4)about the user’s aspiration of category b : number of job applications from user u in category b in the last week, last month, last two months, and last three months.
In the Pull-Based Scenario
• BasicModel
• Basic+TranProb
• Basic+TranProb+Tenure
• Basic+TranProb+TenureProb
CONCLUSION
Q: When is the right time to make a job recommendation and how do we
use this inference to improve the utility of a job recommender system?
the hierarchical proportional hazards model
real-world job application data : Linkedin
Is This App Safe for Children? A
Comparison Study
of Maturity Ratings on Android
and iOS Applications
ABSTRACT
we develop mechanisms to verify the maturity ratings of mobile apps and investigate
possible reasons behind the incorrect ratings.
INTRODUCTION
Android maturity rating policy
“Everyone,” “Low Maturity,” “Medium Maturity,” and “High Maturity,”
iOS’s policy
“4+,” “9+,” “12+,” and “17+.”
iOS rates each app submitted according to its own policies
Android apps are purely a result of app developers’ self-report.
INTRODUCTION
Android rating policy is unclear, and it is difficult for developers to understand the
difference between the four maturity-rating levels
Contribution:
We develop a text mining algorithm to automatically predict apps’ actual maturity ratings from
app descriptions and user reviews.
By comparing Android ratings with iOS ratings, we illustrate the percentage of Android apps with
incorrect maturity ratings and examine the types of apps which tend to be misclassified.
We conduct some preliminary analyses to explore the factors that may lead to untruthful
maturity ratings in Android apps.
RESEARCH QUESTIONS
Does iOS rating strictly reflect its policy?
Are app ratings reflected in app descriptions and user reviews? If so, can
we build an effective text mining approach to predict the true rating of an
app?
Do Android developers provide accurate maturity ratings for their own
apps? For apps published in both markets, are Android ratings consistent
with iOS ratings?
What are the factors that could lead to untruthful maturity ratings in
Android apps in comparison to iOS apps?
discrepancies
• Android does not consider horror content ( C ) as mature content, while iOS does
include
• Android considers graphic violence ( B3 ) as mature content while iOS directly rejects
apps with graphic violence.
• Android integrates privacy protection in its maturity rating policy by including the
social feature ( I ) and location collection ( J ). However, no corresponding privacy-
related consideration exists in the maturity rating scheme by iOS.
• Frequent/intense cartoon violence and fantasy violence ( A2 ) is rated as “Medium
Maturity” (i.e., level 3) in Android but as “9+” (i.e., level 2) in iOS.
• Frequent/intense simulated gambling ( H2 ) is rated as “High Maturity” (i.e., level 4) in
Android but is rates as “12+” (i.e., level 3) in iOS.
we can now use iOS actual maturity rating as a
baseline to examine the reliability of Android apps’
maturity ratings.
Comparing Apps on iOS and Android
For each Android app, we choose
up to 150 search results from the iOS
App Store. For those showing similar
app names, we conducted analysis
to determine the closest fit.
their descriptions and developers’
company names
apps’ icons and screenshots
ALM—Automatic Label of Maturity
Ratings for Mobile Apps
“Android-only” apps
ALM is a semi-supervised learning algorithm, and it processes apps’ descriptions and
user reviews to determine maturity ratings.
1. Building seed-lexicons for objectionable content detection
2. Assigning initial weights to seed-terms
3. Classification
4. Expanding seed-lexicons and adjusting weights
1.Building seed-lexicons for
objectionable content detection
Apps are organized based on their rating scheme together with their
corresponding token, such as A1.txt , A2.txt , B1.txt , and H2.txt .
Human experts read grouped app descriptions and select seed
lexicons to detect objectionable content.
grouped into three bigger lexicons denoted as Ti, i∈ 9,12,17 for classifying
the maturity rating: 9+, 12+, and 17+
2.Assigning initial weights to seed-
terms
Pi, Ni
For each seed-term t, denote its frequency in Pi and Ni as tp and tn
3.Classification
For each app 4 , all terms in its description are selected and categorized
as a set A=tk
maturity rating ma
EXPERIMENT
A total of 1,464 apps were found on iOS App Store and the rest 3,595 apps were
classified as Android-only apps.
Overrated Android Applications
possible reasons
Intelligence
Simulated Gambling
Violence
Mature and Suggestive Themes
Experiment 3: Exploring Factors
Contributing to Incorrect Ratings
apps’ attributes :
popularity, price, and dangerous level of the required permissions .
Developers’ attributes : general privacy awareness, trustworthiness, actual privacy
awareness, and child safety awareness .