Www 2008/ Refereed Track: Rich Media April 21-25, 2008. Beijing, China User-defined Tags Candidate Tags Recommended Tags Sagrada Familia Sagrada Familia: Barcelona Gaudi Catalunya d Spain architecture architecture 正| church church 8 Barcelona G0 au6 d i Catalunya Euro Figure 4: System overview of the tag recommendation process. Vote. The voting strategy computes a score for each candi- Stability-promotion. Considered that user-defined late tag cEC, where a vote for c is cast, whenever cE Cu ags with very low collection frequency are less reliable than tags with higher collection frequency, we want to promote those tags for which the statistics are more stable. This is achieved with the following function: stability(u) kis +abs(ks -log(luD)) computed as: In principle this is a weighting function that weights score(c): =>vote(u,c), (4) he impact of the candidate tags for a given user- u∈U defined tag. lul is the collection frequency of the tag u and ks is a parameter in this function, which is de- ermined by training. The function abs(a)returns the Sum. The didate tag list of the tags lists(C), and sums also takes the union of all can- absolute value of a as over the co-occurrence valt f a candidate tag c E C as cal- Tags with very high culated as: frequency are likely to be too general for individual score():=∑(P(c),ifc∈Cn) hotos. We want to promote the descriptiveness b damping the contribution of candidate tags with a very high-frequency The function P(clu) calculates the asymmetric co-occurrence value, as defined in Equation 2. Note that the score of candi- descriptive(c) g(lcD)) date tag c is obtained by only summing over the tags cE Ct We will use these two aggregation strategies as This is another weighting function, now only applied to line for our evaluation as is presented in Section 6 re-value the weight of a candidate tag. kd is parameter in this function, and is configured by training Promotion. In Section 3 vations with respect to tagging behaviour. In this section Rank- promotion. The co-occurrence values of tag we translate these observations into a"promotion function provide good estimates of the relevance of a candi- to promote more descriptive tags for recommendation. date tag for a user-defined tag. In principle, this is From the tag frequency distribution presented in Figure 1 already used by the aggregation strategy for summing, we learnt that both the head and the tail of the power law but we observed that the co-occurrence values decline would probably not contain good tags for recommendation very fast. The rank promotion does not look at the co Tags in the tail were judged to be unstable descriptors, due ccurrence value, but at the position r of the candidate to their infrequent nature. The head on the other han tag c E Cu for a given user-defined tag u contained tags that would be too generic to be useful(2006 2005, wedding, etc. rank(u, c) k+(r-1Recommended Tags Gaudi Spain Catalunya architecture church Candidate Tags Sagrada Familia: Barcelona Gaudi Spain architecture Catalunya church Barcelona: Spain Gaudi 2006 Catalunya Europe travel User-defined Tags Sagrada Familia Barcelona Tag Co-occurence Tag Aggregation & Ranking Figure 4: System overview of the tag recommendation process. Vote. The voting strategy computes a score for each candi￾date tag c ∈ C, where a vote for c is cast, whenever c ∈ Cu. vote(u, c) =  1 if c ∈ Cu 0 otherwise (3) A list of recommended tags R is obtained by sorting the candidate tags on the number of votes. A score is therefore computed as: score(c) := X u∈U vote(u, c), (4) Sum. The summing strategy also takes the union of all can￾didate tag lists (C), and sums over the co-occurrence values of the tags, thus the score of a candidate tag c ∈ C as cal￾culated as: score(c) := X u∈U (P(c|u) , if c ∈ Cu) (5) The function P(c|u) calculates the asymmetric co-occurrence value, as defined in Equation 2. Note that the score of candi￾date tag c is obtained by only summing over the tags c ∈ Cu. We will use these two aggregation strategies as the base￾line for our evaluation as is presented in Section 6. Promotion. In Section 3 we have made a number of obser￾vations with respect to tagging behaviour. In this section, we translate these observations into a “promotion function” to promote more descriptive tags for recommendation. From the tag frequency distribution presented in Figure 1, we learnt that both the head and the tail of the power law would probably not contain good tags for recommendation. Tags in the tail were judged to be unstable descriptors, due to their infrequent nature. The head on the other hand contained tags that would be too generic to be useful (2006, 2005, wedding, etc.). • Stability-promotion. Considered that user-defined tags with very low collection frequency are less reliable than tags with higher collection frequency, we want to promote those tags for which the statistics are more stable. This is achieved with the following function: stability(u) := ks ks + abs(ks − log(|u|)) (6) In principle this is a weighting function that weights the impact of the candidate tags for a given user￾defined tag. |u| is the collection frequency of the tag u and ks is a parameter in this function, which is de￾termined by training. The function abs(x) returns the absolute value of x. • Descriptiveness-promotion. Tags with very high frequency are likely to be too general for individual photos. We want to promote the descriptiveness by damping the contribution of candidate tags with a very high-frequency: descriptive(c) := kd kd + abs(kd − log(|c|)) (7) This is another weighting function, now only applied to re-value the weight of a candidate tag. kd is parameter in this function, and is configured by training. • Rank-promotion. The co-occurrence values of tags provide good estimates of the relevance of a candi￾date tag for a user-defined tag. In principle, this is already used by the aggregation strategy for summing, but we observed that the co-occurrence values decline very fast. The rank promotion does not look at the co￾occurrence value, but at the position r of the candidate tag c ∈ Cu for a given user-defined tag u: rank(u, c) = kr kr + (r − 1) (8) 331 WWW 2008 / Refereed Track: Rich Media April 21-25, 2008. Beijing, China