Next generation linked in talent search

Editor's Notes

• #4: LinkedIn is the world’s largest professional network, with more than 400M members from all over the world. Its mission: “To connect the world's professionals to make them more productive and successful”. Our member base is growing fast with more than 2 members joining the network every second.
• #5: Our vision is to create the first economic graph digitally mapping the global economy to connect talent with opportunity. In this vision, search plays a big role. On one hand, search helps members to discover opportunity available to them. On the other hand, companies could use search to find the right talent for them.Thus, search plays a central role at LinkedIn. There is a huge variety in profiles found on LinkedIn and we have identified millions of companies, thousands of titles and thousands of skills that can be used to locate a professional that one may intend to hire.
• #7: Given the role of LinkedIn search, we provide multiple search products specializing for different use-cases. The first product is flagship search, which is shown on the left side. This is open to all of our members to find job opportunities, to search for people to connect with or to discover content on LinkedIn to develop their careers. We also offer premium products like Sales Navigator and Recruiter Search. Sales Navigator is to help members discover sales opportunities and help them connect with the opportunities Recruiter search is to help corporation around the world find the right talent for them.
• #9: Search on LinkedIn has two aspects: find and be found. The former is how the system retrieves and ranks results given a query. The later is how our members show case their professional expertise in order to be found. On LinkedIn a typical way to do that is via skills. Here are typical skills of bigdata engineer. There are about 40K std skills on the site. As you can see in the picture, members can endorse their connections on some certain skills. So, if you get endorsement from your peers or managers or from experts on the skills, that’s a strong signal to represent your professional expertise. Given the skills on our eco-system, one of the unique value propositions of LinkedIn is the ability to search by skills.
• #10: The problem of expertise search on LinkedIn has a set of unique challenges. The first is scale. We have 450MM x 35K, which is by far bigger than expert finding task in the previous research, e.g., the task in TREC. Thus, it’s challenging to infer expertise level of each member of every skill he or she might have The second issue that a member typically doesn’t list all of the skills they know on profile ... It is crucial to infer expertise on the skills that members might have but don’t explicitly list Finally, Linkedin search is highly personalized. For instance, … searchers do not only care of expertise of results. So, the search ranking function has to combine those things together.
• #11: In our case, the decision maker is a searcher with a job position in mind looking for candidates. The record of each candidate is his professional career represented by a LinkeIn profile, including … Reputation reflects the level of expertise of the candidate on certain skill. Eventually, similar FICO scores, which is a source that lenders look at to make a judgment whether to give a loan, our goal for skill reputation scores is one the factors (not the only one) that the searcher counts on to make a judgment on whether to hire the candidate.
• #12: Given the goal, here is our approach. In the first step, we use a supervised learning algorithm combining a bunch of signals on the site such as … to infer scores for a subset of (member, skill) pairs. The score means … At this step, the matrix is very sparse since we can be quite certain only for a small percentage of the pairs.
• #13: Once we have the initial matrix, we factorize it. This steps is quite similar to collaborative filtering and LSI, for those of you who are familiar with them. Our member-skill matrix is similar to Netfllix movie rating matrix or document-term matrix in LSI, where a cell is frequency of a term in a document. After factorization, each row in the member matrix could be understood as an unnormalized distribution over latent topics, where a latent topic is a cluster of related skills. Similarly, a column in skill matrix represent a skill in the latent topic space.
• #14: Then, we compute dot-product between member latent matrix and skill latent matrix to reconstruct the original member-skill matrix as well as fill in the unknown cells Take skill co-occurrence patterns to infer missing skills: members knowing “Big Data” are also likely to know “Hadoop”
• #16: LinkedIn maintains the professional profile for over 300 million members. With such a large number of profiles, high-quality profile discovery becomes a challenging problem. To solve the problem, we developed Similar Profiles recommender , which is a fundemantal technology of linkedin’s people recommendations systems. In addition to serving as an independent product, it also powers various profile discovery systems in Linkedin, for example, People You May Want to Hire is a personalized candidate discovery engine for recruiters that takes into account all of the context with regards to recruiting activities Similar Profiles Recommender System helps recruiters and hiring managers discover other similar talent by pivoting of a model user profile.
• #17: Representation of profiles: IN current similar profile recommender, one profile is modeled as a set of fields, such as tile, specialty and company. For each field, we extract the keywords and create a vector of keywords to represent the field Obviously , this model is not able to capture temporal information, not to mention to represent a career path. A natual way to solve this problem is to keep the sequence of work positions. And for each position, we can describe its different dimensions such as The timeline is a sequence of work positions for people’s. Each position is a particular job experience, containing time duration, title information, along with key words summary.
• #18: No need to explain the solution. Just focus on the problem. We use keywords based matching methods to find similarity scores for field-pairs. An overall similarity score is calculated using a weighted linear combination of these scores. Then this overall score is normalized using a logit function, so that it can be interpreted as a probability of being similar. Here, weights reflect the relative importance of field pairs that are matched. These weights are learned by fitting a logistic regression model on train- ing data obtained from active recruiter usage of Similar Profiles product.
• #19: Considering each career path can also be modeled as a set of position, we can find optimal match of pairs of positions from two sets.
• #20: We first point out the different between the set based and sequence base. Generally speaking, given two sequences of nodes (profiles), we conduct sequence alignment to calculate the similarity between them Each node here is a representation of one par- ticular work experience. In sequence alignment algorithm [16], the sequence level sim- ilarity is measured by calculating the sum of the optimal align- ment of node pairs.
• #22: Search by Ideal Candidate: Retains the ability to switch between the ideal candidate and the search by criteria flows Ability to abstract intent from the list of candidates Uses latent features to automatically expand the list of companies and list of industries and skills Takes Skill reputation and title correlation into account.