Recommendation System-Diagram Method

The following is a list of experiment articles by referring to Zhou Tao's two-part graph articles:

1, 2007pre bipartite network projection and personal recommendation.pdf (network structure)

2, 2010pnas-solving the apparent diversity-accuracy dilemma of recommender systems.pdf (material diffusion and Heat Conduction)

Simple Fusion Algorithm of heats and pros

3, 2009njp accurate and diverse recommendations via eliminating redundant correlations.pdf (multi-step material diffusion)

2008epl effect of initial configuration on network-based recommendation.pdf (initial resource allocation)

These two articles are improvements to the pros algorithm, which are significantly improved by 9%, respectively (in my test, the accuracy rate ranges from 16% to 17.8% ), 20% (in the test, the value ranges from 17.8% to 20.5%, and the value can be greater than 21%)

In addition. In this series of articles, some new discriminant indicators are proposed, which are different from the precision, recall, popular, and cover (coverage rate) evaluated as usual (see the recommendation system practice)

The following discriminant indicators are used:

Rangking score: in fact, it is the first place to recommend a correct item, for example, 30 th. If there are 1500 items not collected, rat = 0.02%. The smaller the value, the better the result.

Novelty = mean (log2 (U/Ka), U is the total number of users, and Ka is the number of users of select item A. In fact, it is the reciprocal log of Pop. The larger the value, the better.

Diversity: I think this is a very creative indicator. The author divides it into intra-similarity and inter-diversity)

1. Intra: the similarity between different items of the same user should be smaller. IL = 1/(L L-1) * Σ I + jsij, L is the length of the recommendation list, I = 1/M * Σ 1 ~ M I, m is the number of users

SIJ is obtained by Cf. SIJ = (k (OI) * K (OJ)-1/2 * Σ 1 ~ M (AIL * ajl)

2. Inter: hij = 1-qij (N)/n = 1-Q/l. Q is the user I, and the J recommendation list contains the same number of items.

S = 1/(m-1) Σ hij. Obviously, the larger the inter-diversity, the better.

There is no direct relationship between I and S.

Data sets with Movielens-100K, indicators are using the traditional precision, as well as recall, Pop, etc., can not

The following is the experiment result:

1. Resource initial allocation parameter Beta-precision

2, Beta =-0.3, W + Alpha * w ^ 2, multi-step Diffusion, eliminating redundant information, alpha-precision

 

3, alpha =-0.75, beta-precision

From the result chart, it is better to determine Alpha and then determine the beta results. It is not clear whether there is a better way to adjust these two parameters at the same time to optimize the results, in this paper, Beta =-0.8, alpha =-0.75, but because the indicators here are different from them, the parameters may be slightly different, the result of 21% is better than the result of 50 in the user-KNN algorithm. The memory usage is almost the same. In time, this method only needs 3 ~ 5S, and KNN-50 to 80 s or so.

Recommendation System-Diagram Method