1、Building A Performant Hashtag Ecosystem For LinkedIn2014 LinkedIn Corporation. All RightsReserved.Our mission is to connect the worldsprofessionals to make them moreproductive and successful.Jeff Weiner (LinkedIn CEO)3Agenda Flagship Strategy & How Does Hashtags Help Project Impact Product Flows Has
2、htag creation, consumption and follow The Hashtag Ecosystem Key TakeawayStrategy & ImpactHow Does Hashtags HelpLinkedIn Flagship StrategyFlagships objective is to enablemembers to build and interactwith an active professionalcommunity to advance theircareer (stay informed, buildmeaningful relationsh
3、ips, get ajob)As we transform LinkedIn to aplace where members formcommunities around the thingsthat matter most them, hashtagsare foundational in enablingmembers to express what theywant to talk about and find otherswho share the same interests6Project Impact Scale our hashtag ecosystem made positi
4、ve long lasting impact to feed engagement.(+0.11% Total Homepage Uniques, +0.78% Engaged Feed Users, +0.68% Feed ViralAction Uniques) Enhance our hashtag ecosystem brought hashtag awareness to members and positivelyshifted member posting behavior to post more with hashtags without droppingcreation v
5、olume. (Neutral for Net Content Shared, +81.85% Total Member OnsiteHashtag Posts)2014 LinkedIn Corporation. All RightsReserved.Product FlowsHashtag In Post Creation Flow Typeahead is a great way ofrecommending relevant hashtagsfor user to leverage when creating apost Adding hashtags when creating po
6、stallows a member to explicitlyexpress his intentionHashtag Consumption - Search On consumption side, hashtagsfacilitate content discovery andorganization e.g. #LinkedIn,#MachineLearning, #AI All members posts includinghashtags are indexed by LinkedIn,thereby set the foundation forhashtag consumptio
7、n experienceHashtag Consumption - Hashtag Feed A hashtag feed is created bysearching all member postscontaining a given hashtag Search is our funnel, click onhashtags are just shortcut forcontent search Content search and hashtagfeed are two sides of thesame coinHashtag Follow“Because you follows .”
8、 Update In FeedWhat is this?Once a member follows a hashtag, they will startseeing relevant posts contain that hashtags in theirmain feedHow is this generated?The content shown is this update is editoriallyselected currently, with algorithmically taggedcoming soonHashtag EcosystemHashtag Typeahead R
9、ecommendationHashtag Typeahead RecommendationHashtag Sources & Selection Hashtag recommendations comes from awhitelist of hashtags computed offline. The goal of candidate selection is to select asubset of hashtags in the whitelist which maybe relevant to the post and/or member.Hashtag Typeahead Reco
10、mmendationHashtag Candidate Selection Candidate selection maximizes recall, whilereturning a set of candidates which is asmall enough to be ranked online The current state efficiently filters the entirewhitelist of candidates and ranks by a pre-computed hashtag-engagement based score.Hashtag Typeahe
11、ad RecommendationRanking Ranking of hashtag recommendationsis the core relevance area. Here we begin with a standard MLpipeline which combines member,post, and hashtag features in alogistic regression model Sort by probability of selecting ahashtag : P(h | m, p)Hashtag Feed Serving StackHashtag Feed
12、 Serving DetailsHashtag Service & ScalingHashtag Service Hashtag Service: REST APIs on top of thehashtag SOT store (aka HashtagDictionary) The REST APIs maps a hashtag ID to afull hashtag string Hashtag service also providesintegration with UCF (user contentfiltering) serviceHashtag Service Scaling
13、Hashtag Dictionary is maintainedinside Espresso (LinkedIns NoSQLDocument Store) Since some hashtags tend to becomea lot more popular than others (e.g.#innovation vs. #random), “hotkey”problem occurs Solution: added a Couchbase cachinglayer to serve read request, whichsignificantly reduce read QPS on
14、Espresso nodeHashtag Spam Filtering & Quality ControlHashtag Spam Filtering & Quality ControlUCF Integration:Async UCF Upon hashtag creation, async UCFflow will be triggered whichgenerate UCF records When a hashtag is marked asspam by a member, the async UCFmodule will get notified via akafka event
15、and will update theCCR of that hashtag accordinglyHashtag Spam Filtering & Quality ControlUCF Integration:Consumption (Decoration) When a hashtag Get/batchGetrequest is received, hashtagservice will check CCR to see ifspam is detected. If not, a hashtagid is resolved into a full hashtagstringHashtag
16、 i18n & PersonalizationHashtag i18n & PersonalizationWhat is “Hashtag i18n & Personalization” ?To tag articles in international languages intointernational hashtag feeds where liquidity is low /non existentWhy is this needed?To increase the liquidity of hashtag feed, and ramphashtag globally, we nee
17、d to return contentmatching members language preferenceHashtag i18n & PersonalizationSolution Strategy:Take into account a members understoodlanguage when ranking content within hashtagfeedsImplementation:Associate each member post with a detectedlanguage using Microsoft Text Analytical ServiceAPIEn
18、hance the search stack to query based on thelanguage label and members understoodlanguageSummaryKey Learnings From Building The Hashtag Ecosystem: Start from a coherent product experience that is well aligned with businessstrategy Follow up with measurement: success is defined as when we seecommunities gathering and conversations happening around hashtags,worldwide Doubling down on personalization, scalability, content quality andinclusiveness from engineering side to deliver our goalAppendix31Hashtag Ecosystem ArchitectureThank You33
