[paper reading] ckan: collaborative knowledge aware autonomous network for adviser systems

original text ：CKAN
Code ：https://github.com/weberrr/CKAN

What: Put forward CKAN, One will Cooperative signal And Knowledge is connected The method of natural combination .
Why: Existing methods only focus on KG Medium Knowledge is connected （knowledge associations）, Ignored Cooperative signal （ collaborative signals）, This is often user-item Lacking in interaction .
How: Propose heterogeneous communication strategies , Encode two kinds of information naturally , Then the attention mechanism of knowledge perception is applied to distinguish the contributions of different knowledge-based neighbors .
Result:
Conclusion:

CKAN The previous method was not considered at the same time 【 Knowledge is connected 】 and 【 Cooperative signal 】 These two kinds of heterogeneous information , So try to combine the two kinds of information by natural coding .

Model

1. Heterogeneous Propagation

1.1 Collaborative communication

KG chart ：

item-entity chart ：

collaboration propagation

【user By interactive item To represent 、item By interactive user Interactive item To represent 】

【 The initial user said （initial entity set of user）】：（ from user Interactive item To express ）

among Yuv=1 Represent user u And objects v There is interaction ;e It's an entity ,(v,e) Indicates an item v And objects e There's a connection .
It means ,【 user 】 It is interacted by item, these item Related entities entity To represent the .【 Express with entity 】

Multiple users interact with the same item , Between them is 【 Users' collaborative neighbors 】.
Multiple items are interacted by the same user , Between them is 【 Collaborative neighbors of items 】.

Collaborative neighbors of items ：

Vu Refers to the collaborative neighbors of items ,vu finger Vu This set Items in . Collaborative neighbors of items , namely item1,item2 All be user1 Interaction ,item2 yes item1 Objects of collaboration neighbors . Use entities related to collaborative neighbors entity To represent the .【 Express with entity 】

【 The initial item represents （initial entity set of item）】：（ from item Interactive user Interaction item To express ）

The objects are represented by their cooperative neighbors and their associated entities .

1.2 Knowledge map dissemination

Therefore, we can get after multi-level recursion user and item It means ,l-th Tail entity of , Aggregated (l-1)-th Head entity h Information about .
（user and item Multi level recursion Entity representation A general representation of ）

Empathy , Get multi-level recursion Triples represent A general representation of ：

【the knowledgebased high-order interaction information of user and item is successfully captured】

2. Knowledge-aware Attentive Embedding

The previous operation is represented by the entity of the object user, Express with the entity of neighbor items item, Collaborative communication is considered ; Then through multi-layer recursive propagation , Knowledge map propagation is considered , Get the fused head entity h And different relationships r Tail entity of t, But we also want to further distinguish the tail entities t The difference between , So add Attention mechanism To distinguish the tail entities with multi-layer information t.

Tailstock t Attention embeddedness indicates ：（ By the head entity h And relationship r control ）

among ,


Call it attention factor , Its calculation process is as follows ：

And pass softmax Normalize ：

So you get l-th Representation of triples of ：（ It is divided into user and item, Here is the general expression ）

You can see the calculation l-th When represented by triples of , Calculate the attention of each triplet in this layer , It is equivalent to aggregating all triples of this layer .

For the initial layer （ The first 0 layer ）, hold user and item Corresponding entity Directly add to represent ：

The author believes that the most central place is item Those entities that are directly related , And item It is closest in the buried layer , So add them to , To represent the origin：

So you get user and item Of additional representations：

3. Model Prediction

Each layer emphasizes different high-order connectivity and user preferences , Therefore, for the above 【user and item Of additional representations】 Aggregate separately , Adopted 3 Different ways ：
Sum aggregator：

Pooling aggregator：

Concat aggregator：

Recently, I have arrived at the aggregation user Represents and item Indicates inner product ：

Reference resources ：

1. CKAN Paper notes ：https://zhuanlan.zhihu.com/p/181475023

attach ：
Two pictures