A Scalable, High-performance Algorithm for Hybrid Job Recommendations

Toon De Pessemier, Kris Vanhecke, Luc Martens,
September, 2016
iMinds – Ghent University, Belgium
toon.depessemier@ugent.be
A Scalable, High-performance Algorithm
for Hybrid Job Recommendations

A Scalable, High-performance Algorithm for Hybrid Job Recommendations
Toon De Pessemier, Kris Vanhecke, Luc Martens
2
Introduction: Job recommendations
Not a classic recommender story
Not a classic solution
 Specific metadata characteristics
 Discipline, industry, career level, …
 Detailed user profile
 Experience, education (university degree), employment
 Limited availability in time (active_during_test)
 Various user-item interactions
 Click, bookmark, reply, delete
 Specific meaning of delete (click on “X”  load new item)
 Impressions
 Recommendations generated by XING’s recommender  Bias

3
Our goals
 XING’s evaluation measure
 Reflects typical XING use case
 Scalable
 Number of users and items
 Dataset = subset of XING users
 Incremental updates
 Continuous stream of new job items
 Updating models instead of recalculating
 Fast score calculation
 New job items  fast distribution to target users
 Limited computational resources

4
Findings
 Challenge = Prediction task
 ≠ Recommendation task
 No influence on user behavior
 Recommendations are not evaluated
by the user
 Important quality metrics are not evaluated
 Usefulness
Risk: Items already discovered by the user
Items that the user already interacted with, can be recommended
 Diversity
Risk: Too much of the same
 Serendipity
Risk: Items that are difficult to find but interesting, are unfairly evaluated as
“poor recommendations”

5
Findings
 The information value of impressions is
limited
 Recommendations of existing job
recommender
 Bias to Xing’s algorithm
 Less diverse
 Subset of recommendations
 No guarantee that the user has seen the item
 No cold start user  Better results if only the
interactions are used
 Penalty for items with a limited visibility
 Low visibility  low probability of interaction
 Low visibility  penalty  better results
 Item visibility estimated by number of interactions in training
set

6
Findings
 Influence of the user’s region
 Expected: interest for jobs located in the user’s
home region or in adjacent regions
 Observed: Many interactions for jobs located in
non-adjacent or far away regions
 E.g. Users of Lower Saxony  Jobs in Baden-
Württemberg
 Many cold-start users
 No interactions, no impressions (9.7%)
 CB recommendation based on explicit profile
 Risk: too general or to specific profile
 Risk: not updated by the user

7
Findings
 Traditional classification does not work
 Positive class: click, bookmark, reply
 Negative class: delete
 Recommendations: items most typical for the positive class
 Poor score
 Reasoning: meaning of delete action
 Click on X button in recommendation list
 New recommendation will be loaded and displayed
 Deletes not sampled from complete job offer but from
recommendations (bias: items more similar to the user’s interests
than random items)
 Not necessarily a disinterest of the user
 Intension to click: new recommendation

8
Content-Based Recommender
 Based on feature matching
 Explicit user profile
 Interactions  counter for each feature
 Interaction weight
 Updating counters
 Delete=0, click=1, bookmark=10, reply=10 (no significant effect of deletes)
 Positive counters (posf,u)  item has feature
 Negative counters (negf,u)  item does not have feature
 Score calculation
 α = 0.5 (positive counters are more important than negative counters)
 IDF = inverse document frequency: feature frequency across all jobs
 N = total number of items
 nf = number of items with feature f
 wf = weight per feature type (tag, discipline, industry, …)
 u = user
 i = item
score(u,i) =
1
𝑓𝜖 𝑖
𝑓∈𝑖
𝑤𝑓 𝑝𝑜𝑠 𝑓,𝑢 − 𝛼 𝑛𝑒𝑔 𝑓,𝑢 𝑙𝑜𝑔
𝑁
𝑛 𝑓

9
Content-based calculation
 Profile
 Offline calculation
 Incremental updates of counters
 IDF
 Slightly varying over time
 Periodic updates
 Target items
 Active items
 Minimum matching threshold (positive counters and item
have X features in common)
 Algorithm running in parallel for different users
 Fast calculation of the recommendations

10
Collaborative filtering: KNN
 Traditional KNN
 Distance based on interactions
 Our KNN solution
 Distanced based on interactions and metadata
 2 items are similar if users have interacted with both
 2 items are similar if they have metadata features in common
Feature distance: factor 𝑙𝑜𝑔
𝑁
𝑛 𝑓
 Fine-grained distance function
 Risk of ties is reduced
 Method:
 For each candidate item:
 Calculate distance to k-nearest items that the user has positively interacted with
 Select items with shortest distance
 𝑠𝑐𝑜𝑟𝑒 𝑢, 𝑖 =
1
𝑘 𝑘
𝐷𝑖𝑠𝑡 𝑚𝑎𝑥−𝐷𝑖𝑠𝑡 𝑖,𝑘
𝐷𝑖𝑠𝑡 𝑚𝑎𝑥
 Based on Weka Framework
 BallTree implementation of NearestNeighbourSearch package

11
KNN calculation
 Item distances
 Offline calculation
 Slightly varying over time
 If partially computed distance > threshold
 stop calculation
 Score calculation
 Fast if distances are precomputed
 Algorithm running in parallel for different users
 Fast calculation of the recommendations

12
Results and fallback
 CB: 286,041.10
 KNN: 298,316.85
 Hybrid: 344,264.37
 Fallback cold start users:
 No interactions:
 KNN based on interactions is not possible (26.5% of users)
 No interactions  use impressions (16.8% of users)
 Solution without fallback to impressions (only based on profile):
292,909.26
 No interactions and no impressions (9.7% of the users):
 Hybrid  CB
 CB cannot generate recommendations:
 For 1485 users
 Recommend the 30 most popular items (most positive interactions)
 Without fallback to most popular recommender: 344,241.51
 Most popular recommender as the only solution: 73,298.13

13
Questions?

A Scalable, High-performance Algorithm for Hybrid Job Recommendations

More Related Content

What's hot (20)

Similar to A Scalable, High-performance Algorithm for Hybrid Job Recommendations (20)

Recently uploaded (20)

A Scalable, High-performance Algorithm for Hybrid Job Recommendations