PDF Publication Title:
Text from PDF Page: 007
Practical Diversified Recommendations on YouTube with Determinantal Point Processes CIKM’18, October 2018, Turin, Italy Algorithm 1 Rank a video feed via a DPP Require: α,σ Strategy Fuzzy deduping Sliding window Smooth score penalty DPPs Deep DPPs Satisfied homepage watchers -0.05% -0.26% -0.41% +0.63% +1.72% ▷DPPparameters ▷ The size of the windows ▷ The quality estimates and video embeddings Require: W ←{1,2,...,N} ▷Indicesoftheitemstorank Require: k ∈ N Require: q, φ Construct D from φ L ← N × N -matrix for i = 1, 2, . . . , N do ▷ Compute distances from embeddings ▷ Construct the DPP Table 1: Experimental results over approximately 1 week for various attempts at improving video diversity in users’ feeds. “Satisfied homepage watchers” refers to a metric that consid- ers sessions originating from the homepage and counts how many of these sessions were of significant duration—a no- tion of homepage utility. Only DPPs improved this metric. to videos 1 through n that have already been selected: qnew,v = qoriginal,v ∗ e−b(φv ·φprevious) (16) n with φprevious = an−k−1φk , (17) k=0 where q is the quality score we sort by, a and b are free parameters, and φ is the embedding vector. As seen in Table 1, all of these attempts led to a less useful mobile homepage feed, as measured by the number of users with long sessions originating from homepage. When experimenting with DPPs, we first used the kernel L de- scribed in Section 4.2, and evaluated a variety of embeddings and dis- tance functions (dense and sparse audio embeddings, frame embed- dings, thumbnail image embeddings, document text embeddings, etc.). We found that it works quite well to use Jaccard distances for Di j in Equation 10, applied to sparse vectors φ consisting of item tokens. (For example, the Saturday Night Live video “Olive Gar- den -SNL” has tokens “snl”, “olive garden”, “saturday night”, “night live”, and “sketch”, among others.) Live experiments on YouTube’s mobile homepage recommendations saw dramatic improvements for our users. In addition to the +0.63% on the satisfied homepage watchers metric shown in Table 1, we also saw +0.52% in overall watch time, which is quite a significant jump over the baseline. Because of this success on mobile, diversification via DPPs has been deployed on all surfaces, including TV, desktop, and Live streams. (Note that while the deep Gramian DPPs system looks very promis- ing on the “satisfied homepage watchers” metric, it has not yet been deployed. As mentioned earlier, these deeper models change the ranking substantially enough from the un-diversified baseline that secondary business metrics begin to be significantly impacted, requiring additional tuning.) Interestingly, for some choices of parameters we saw losses in direct interactions on the homepage, though across the site we had an overall win. Figure 5 shows the percent increase in view time that originates from the homepage. This suggests that users find content sufficiently attractive that it leads to longer sessions starting from the homepage. And indeed, we did observe increased activity on the related videos panel (a panel of videos one sees alongside the video that is currently playing), in terms of click-through rate, number of views, and amount of total view time, despite the fact that our original change only affected the videos shown on the homepage for j = 1, 2, . . . , N do Lij ←LW[i]W[j](α,σ,q,D) end for end for R ← [] while |W | > 0 do ▷ The final ordering M ← GreedyApproxMax(L, min(k, |W |)) D←∅ for i ∈ M do ▷ Indices into W R←R+W[i] ▷Getitemsfromthegreedyapprox D ← D + W [i ] end for W ← W \ D L ← L[W ,W ] end while return R problem: ▷ Remove the selected items ▷ Restrict L to the W submatrix max Y:|Y |=k det(LY ) . (14) As shown in [18], this maximization is NP-hard. In practice though, a standard greedy algorithm for submodular maximization from [31] seems to work well for approximately solving this problem. The greedy algorithm starts from Y = ∅ (the empty set), then runs k iterations, adding one video to Y on each iteration. The chosen video in iteration i is the video v that produces the largest determinant value when added to the current chosen set: max det(LY ∪v ) . v ∈remaining videos (15) Beyond its simplicity, an additional advantage of using this greedy algorithm is that, if we keep track of the order in which greedy selects videos, then this gives us a natural order for the videos in the corresponding size-k window of the user’s feed. Algorithm 1 summarizes the ranking algorithm described in this section. As we will see in the subsequent section, this ranking helps users find the content that they want to consume more easily. 5 EXPERIMENTAL RESULTS First, we will describe some basic comparison baselines. Before finally arriving at DPPs, we tried three diversification heuristics: (1) Fuzzy deduping: disallow any video i whose distance to a video j already in the feed is below a threshold τ: Dij < τ. (2) Sliding window: allow at most n out of every m items to be below a distance threshold τ . (3) Smooth score penalty: When selecting the video v for posi- tion n + 1, re-scale the quality scores to account for similarityPDF Image | Practical Diversified Recommendations on YouTube with Determinantal Point Processes
PDF Search Title:
Practical Diversified Recommendations on YouTube with Determinantal Point ProcessesOriginal File Name Searched:
cikm2018.pdfDIY PDF Search: Google It | Yahoo | Bing
Cruise Ship Reviews | Luxury Resort | Jet | Yacht | and Travel Tech More Info
Cruising Review Topics and Articles More Info
Software based on Filemaker for the travel industry More Info
The Burgenstock Resort: Reviews on CruisingReview website... More Info
Resort Reviews: World Class resorts... More Info
The Riffelalp Resort: Reviews on CruisingReview website... More Info
CONTACT TEL: 608-238-6001 Email: greg@cruisingreview.com (Standard Web Page)