You are on page 1of 3

Key words/concepts: Unsupervised machine learning, personalization/ prediction algorithms,

reciprocal recommendation systems/engines, explicit/implicit data inputs.

"Contemporary digital data analytics systems feed on a diet of data produced through human
activity. Through this feeding, machines receive the informational nutrition required for their own
development: to become smarter, more aware of their environment, more responsive and adaptive
in their interactions with people. By eating human data, machines are learning."

Algorithms.

Influenced by the advent of commercial recommender systems, researchers and practitioners tried
to apply to recruiting inappropriate techniques typically designed for the suggestion of products or
content. One job posting is usually meant to hire one individual. This is different from books or
movies, where as long as there is inventory it can be recommended to many users at the same time.
If we over-recommend a job, we could bombard an employer with thousands of applications.

So it's critical to display a job to a selected number of compatible candidates; displaying it too
widely /untargeted will result in dissatisfaction of all parties. In addition, time is an important factor:
both jobs and unemployed people are available for a limited period of time. Personalization on the
candidates side is based on matching data like 1.tags, 2.distance, 3.recency, (and maybe 4.seniority
level). On the employers' side we'll get into personalization when we'll have to 'fill' the PPA jobs with
applicants.

Content-Based recommender systems. The central idea behind content-based suggestion systems is
to match item (job) characteristics, (keywords, domains), with the existing and acknowledged users’
preferences. There are two problems with this approach: 1.People are bad at introducing matching
characteristics (keywords), or the fixed characteristics (domains) are insufficiently precise to match
supply and demand. 2.On mobile users don't want to enter inputs (keywords or multiple selects) -
just wants to scroll or swipe.

Collaborative Filtering recommender systems. On the contrary, the core concept behind
collaborative filtering [CF] techniques - probably the most widely used filtering technique nowadays -
is to draw on an existing user community information and past behaviors in order to anticipate which
product one particular user of the system might like or be interested in. (amazon).

There are two ways in which we can use CF: 1.Similar tags. People with these tags also have these
top tags (top tags can be formed from the last X (ex.100) tags added to any person - in order to take
into account the recency/time of the tags). 2.Similar jobs. People who applied/saved to this job also
applied/saved jobs xyz (active jobs). My main concern with similar tags is the fact that we operate in
a multi-language environment..

Reciprocal Recommender Systems. Conventional one-sided suggestion strategies are less usable in
recruiting, as two key variables differ: [1] The suggested items for employers are people and [2] both
parties’ interests or needs must be satisfied. These “reciprocal”, “bilateral” or “two-sided” algorithms
operate in a very unique recommendation context in which all individual preferences must be
satisfied This bilateral relationships domain must therefore be modelled considering human double-
sided preferences." (tinder).
Data.

The difference between dumb and smart tags (called in literature "explicit / implicit profile
information /feedback") is not that the dumb ones are not useful - only that the smart ones have
multiple 'quality levels': while one targeted job is the same as the next targeted job, a job you've
interacted with can be remove/save/apply from your side, plus interview/save/ remove from the
employer side - it has more data. Maybe interview = 3, apply = 2, save = 1 (same as
search/target/exper) and remove by me -1, remove by employer -2.. (or also -1).

Experience, My job actions, and Employers' actions - all 3 refer to the same data: job title, employer's
name, and "more tags" (keywords in job, and currently 'domains' in experience, but we should
transform them in 'more tags' as well). In addition, using the smart (implicit) tags don't require the
candidate to be very smart with keywords: tags from the job are automatically added to his profile,
with the right multiple (probably the same multiplier apply to all three -title, more, employer- to
keep it simple for now).

Reciprocal recommender systems, more than any other personalization framework, must cope with
the well-known 'cold start' problem. New users must be suggested with other people profiles in
order to respond and interact with the service without further wait. The use of hybrid methods is a
potential approach to this problem. Another answer lies in the possible integration of an
introductory rating system (select few jobs you're targeting) right in/after the sign-up process.

So you're feeding the machine with applies/save/rejects rather than search/etc., but in order to get
applies/saves, you need to already display some quality jobs (appetizer!!) - and here the idea of 'tell
us your targeted jobs' is useful. CV-1st card is also the first step after registration, so if we stick to
'register before seeing jobs' it makes sense to keep 'targeted jobs' in CV-edit. If we want to display
jobs before registration, however, we need to bring 'select few jobs you're targeting' in front (similar
to 'sign up for email alerts?) - they should be before displaying jobs in any case.

Searches and Targeted jobs are pretty much the same thing - they help us (and the employer, if this is
what we display in the resume) understand what the person wants for the future. Can user searches
fill the 'targeted jobs' fields (maybe in case she didn't fill any)? Ex: top searches from the last 20
performed. Also about 'recency in tags', if we make the top tags of every candidate out of the last
100 inputs, we'll constantly have the freshest 'food'.

There is another time-factor: the activity of the candidate. We might need to display different jobs to
candidates than view/login? We're doing it in alerts. Another area to explore if whether search
results are personalized or the same for all - probably the same for all now, to keep it simple, but
definitely using the similar tags/jobs algorithm to always display some results.

More advanced links:

http://www2013.wwwconference.org/companion/p963.pdf

http://www.joics.com/publishedpapers/2011_8_16_4061_4068.pdf

http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.402.1036&rep=rep1&type=pdf

http://sydney.edu.au/engineering/it/research/tr/tr651.pdf

http://users.cs.fiu.edu/~taoli/pub/p35-li-MEET.pdf

http://www.gdeepak.com/thesisme/Applying%20Data%20Mining%20For%20Job
%20Recommendations.pdf

https://pdfs.semanticscholar.org/e2e4/a8e6887ea0e50ebef372f72093a96cd25e0d.pdf

http://www.cse.unsw.edu.au/~bradfordh/papers/pricai14.pdf

https://www.linkedin.com/pulse/implicit-ecosystem-online-dating-reciprocal-engines-quentin-baltus

You might also like