dssm难点介绍以及相关技巧

arXiv:1801.02294v3 [stat.ML] 21 May 2018

Learning Tree-based Deep Model for Recommender Systems

Han Zhu, Xiang Li, Pengye Zhang, Guozheng Li, Jie He, Han Li, Kun Gai

Alibaba Group

{zhuhan.zh,yushi.lx,pengye.zpy,guozheng .lgz,jay.hj,lihan.lh,jingshi.gk}@alibaba-inc.c om

ABSTRACT

Model-based methods for recommender systems have been stud-

ied extensively in recent years. In systems with large corpus, how-

ever, the calculation cost for the learnt model to predict all user-

more expressive interaction forms between user and item features,

e.g., interactions through deep neural networks, bec ause of the cal-

culation cost.

In t his paper, we focus on the problem of introducing arbitrary

advanced models to recommender systems with lar ge corpus. We

propose a novel tree-based metho d which can provide logarithmic

complexity w.r.t. corpus size even with more expressive models

such as deep neural networks. Our main idea is to predict user in-

terests from coarse to fine by traversing tree nodes in a top-down

fashion and making decisions for each user-node pair. We also

tiveness of the proposed method in production environments.

CCS CONCEPTS

• Computing methodologies → Classification and regression

trees; Neural networks; • Information systems → Recom-

mender systems;

KEYWORDS

Tree-based Learning, Recommender Systems, Implicit Feedback

1 INTRODUCTION

Recommendation has been widely used by various kinds of content

providers. Personalized reco mmendation method, base d on t he in-

recommendation algorithms may fail to predict from the entire

corpus. The linear prediction complexity w.r.t. the corpus size is

unacceptable. Deploying such large-scale recommender system re-

quires the amount of calculation to predict for each single user

be limited. And besides preciseness, the novelty of recommended

items should also be responsible for user experience. Results that

only contain homogeneous items with user’s historical behaviors

are not expected.

To reduce the amount of calculation and handle enormous cor-

pus, memory-based collaborative filtering methods are widely de-

scope of candidate set, i.e., not all items but only items similar to

the triggers can be ultimately recommended. This intuitio n pre-

vents the recommender system from jumping out of historical be-

havior to explore potential user interests, which limits the accu racy

of recalled results. And in practice the recommendation novelty is

also criticized. Another way to reduce calculation is making co ar se-

grained recommendation. For example, the system recommends a

small number of item categories for users and picks out all corre-

sponding it ems, with a fol lowing ranking stage. However, for large

corpus, the calculatio n problem is still not solved. If the category

In the literatures of recommender systems, model-based meth-

ods are an active t opic. Models such as matrix factorization (MF)

[19, 30] try to decompose pairwise user-item preferences (e.g., rat-

ings) into u ser and item factors, then reco mmend to each user its

most preferred items. Factorization machine (FM) [28] further pro-

poses a unified model that can mimic different factorization models

with any kind of input data. In some real-world scenarios that have

no explicit preference but only implicit user feedback (e.g., user

behaviors like clicks or pu rchases), Bayesian personalized ranking

[29] gives a solution that formulates the preference in triplets with

tion stage thus is equivalent to retrieve user vector’s nearest neigh-

bors in inner product space. For vector search problem, indices like

hashing or quantization [18] for approximate k-nearest neighbor

(kNN) search can ensure the efficiency of retrieval.

However, the inner product interaction form between user and

item’s vector representations severely limits model’s capability. There

exist many other kinds of more expressive interaction forms, for ex-

ample, cross-product features between user’s historical behaviors

and candidate items are widely used in click-through rate predic-

tion [5]. Recent work [ 13] proposes a neural collaborative fi ltering

erate varying user vectors according to different candidate items.

Beyond the above works, o ther methods like product neural net-

work [27] have also proven the effectiveness of advanced neural

networks. However, as these kinds of models can not be regulated

to inner product form between u ser and item vectors to utilize effi-

cient approximate kNN search, they can not be used to recall candi-

dates in l ar ge-scale recommender systems. How to overcome the

calculation barrier t o make arbitrary advanced neural networks

feasible in large-scale recommendation is a problem.

To address the challenges above, we propose a novel tree-based

of information ubiquitously exists in many domains. For exam-

ple, in E-commerce scenario, iP hone is the fine-grained item while

smartphone is the coarse-grained concept to which iPhone belongs.

The proposed TDM method l everages this hierarchy of informa-

tion and turns recommendation problem into a series of hierarchi-

cal classification problems. By solving the problem from easy to

difficult, TDM can improve both accuracy and efficiency. T he main

contributions of our paper are summarized as follows: