阿里基于树结构的召回体系一共发了三篇paper,这是第二篇,其他两篇论文的阅读笔记见下方链接:
-
基于树的召回框架(一):Learning Tree-based Deep Model for Recommender Systems
-
本文讲解论文的
supplementary files 下载
,包括一些论断的证明过程以及 JTM 的代码实现 -
基于树的召回框架(三):Learning Optimal Tree Models under Beam Search
背景
在论文《Learning Tree-based Deep Model for Recommender Systems》中提出的基于树结构的召回模型,其在优化树结构
T
\mathcal{T}
T
和优化用户兴趣模型
M
\mathcal{M}
M
的两个过程使用的评估指标是不一致的,而本文旨在提出一个训练框架,能够将这两个过程在一个统一的指标下进行优化,从而加快训练过程以及提高召回的效果。本文提出的方法称为 JTM。
方法
与第一篇论文的最大不同之处在于,本文使用的树的层级结构是固定的,文章以完全二叉树为例方便说明算法运行过程,并指出该算法可以拓展到 multi-way tree。假设物品集合为
C
C
C
,该完全二叉树的所有叶子节点的集合为
N
N
N
,则满足
∣
N
∣
=
∣
C
∣
|N|=|C|
∣
N
∣
=
∣
C
∣
。本文中对于树结构的优化,实际上是在优化一个映射函数
π
(
⋅
)
\pi(\cdot)
π
(
⋅
)
,其建立了物品到叶子节点的一对一映射。寻找函数
π
(
⋅
)
\pi(\cdot)
π
(
⋅
)
的过程,实际上等价于寻找带权二部图的最大匹配(相关概念自行搜索),其证明过程如下(来自该论文的补充材料,见本文开始部分):
本文并不直接将物品分配到最后一层的叶子节点中,而是从根节点到叶节点、自上而下、由粗到细地进行分配。算法描述如下:
博主注
:个人感觉该算法描述结合论文的文字解释也仍然不是很直观,因此查看了论文给出的(粗略)实现的代码,个人认为Algorithm 2 可以用如下示意图来理解:
我们以16个物品为例(
∣
C
∣
=
16
|C|=16
∣
C
∣
=
1
6
),即对应了有 5 层的满二叉树,并假设我们取
间隔
d
=
2
d=2
d
=
2
。树在初始化时,所有16个物品都被分配到了根节点(第0层):
在第一轮迭代后(对应Algorithm 2 中的第3行到第9行),所有物品都被分配到了第 2 层,如下图所示:
在第二轮迭代后(对应Algorithm 2 中的第3行到第9行),所有物品都被分配到了第 4 层,如下图所示:
需要注意的是,在每一轮迭代,实际上是对于上一轮已经分配好的那层的所有节点一一进行分配,即已经被分配到某个节点的物品,在后续迭代中只会继续被分配到该节点的后代节点中。例如第一轮迭代,实际上是进行了 1 次匹配,即将根节点中的16个物品和第 2 层中的 4 个节点进行匹配;在第二轮迭代中,实际上是进行了 4 次匹配,即对第 2 层的 4 个节点的后代分别进行了一次匹配:将分配到该节点中的所有物品,分配到该节点的下 2 代的后代(即儿子的儿子)中,注意,第 2 层节点的下 2 代的后代位于第 4 层。
最后要解释的问题就是,Algorithm 2 的第5行,即给定一个物品集合
C
C
C
以及节点集合
N
N
N
,匹配过程是如何进行的。例如,上面图示例子中,第一轮迭代时,我们将
c
1
c
4
c
6
c
15
c_1c_4c_6c_{15}
c
1
c
4
c
6
c
1
5
这4个物品分到第 2 层的第一个节点的依据是什么?正如论文中指出的,“ we use a greedy algorithm with rebalance strategy to solve the sub-problem”:
前面已经证明了这个过程相当于寻找带权二部图的最大匹配的问题,而为了降低求解复杂度,本文采用的是贪婪算法。首先,对于所有的物品-节点对,计算
c
k
c_k
c
k
和
n
m
n_m
n
m
之间的边权重
L
c
k
,
n
m
\mathcal{L}_{c_k,n_m}
L
c
k
,
n
m
,即一共要计算
∣
C
∣
×
∣
N
∣
|C|\times|N|
∣
C
∣
×
∣
N
∣
个权重(对应后文代码中的
get_weights()
,注意,此处的
∣
N
∣
≠
∣
C
∣
|N|\neq |C|
∣
N
∣
=
∣
C
∣
,因为此处的
N
N
N
并非所有叶子节点集合)。对于每个物品
c
k
c_k
c
k
,将其与连接后权重最大的节点
n
m
n_m
n
m
连接起来,然后进行rebalance操作(见
assign_parent()
代码的后半部分)。所谓 rebalance,指的是手动调节一些物品分配到的节点,因为可能有的节点分配不到任何物品,而有的节点分配得到了过多节点(第
d
d
d
层的每个节点最多只能分配到
2
l
m
a
x
2
d
=
2
l
m
a
x
−
d
\frac{2^{l_{max}}}{2^d}=2^{l_{max}-d}
2
d
2
l
m
a
x
=
2
l
m
a
x
−
d
个物品)。具体逻辑见后文代码的
assign_parent()
。
算法实现
节选自论文作者提供的代码实现,更多细节见文章开始的supplementary files链接。
class TreeLearner(object):
def __init__(self, filename):
self.filename = filename
self.tree_meta = tree_proto.TreeMeta()
self.id_code = dict()
self.item_codes = set()
self.nodes = dict()
def get_ancestor(self, code, level):
code_max = 2 ** (level + 1) - 1
while code >= code_max:
code = int((code - 1) / 2)
return code
def get_nodes_given_level(self, level):
code_min = 2 ** level - 1
code_max = 2 ** (level + 1) - 1
res = []
for code in self.nodes.keys():
if code >= code_min and code < code_max:
res.append(code)
return res
def get_children_given_ancestor_and_level(self, ancestor, level):
code_min = 2 ** level - 1
code_max = 2 ** (level + 1) - 1
parent = [ancestor]
res = []
while True:
children = []
for p in parent:
children.extend([2 * p + 1, 2 * p + 2])
if code_min <= children[0] < code_max:
break
parent = children
output = []
for i in res:
if i in self.nodes:
output.append(i)
return output
# .......
def get_itemset_given_ancestor(pi_new, node):
res = []
for ci, code in pi_new.items():
if code == node:
res.append(ci)
return res
def tree_learning(d, tree):
""" The overall tree learning algorithm (Algorithm 2 in the paper)
Returns:
the leant new projection from item to leaf node (\pi_{new})
Args:
d (int, required): the tree learning level gap
tree (tree, required): the old tree (\pi_{old})
"""
l_max = tree.tree_meta.max_level - 1
l = d
pi_new = dict()
# \pi_{new} <- \pi_{old}
for item_code in tree.item_codes:
ci = tree.nodes[item_code].id
pi_new[ci] = tree.get_ancestor(item_code, l - d)
while d > 0:
nodes = tree.get_nodes_given_level(l - d)
for ni in nodes:
C_ni = get_itemset_given_ancestor(pi_new, ni)
pi_star = assign_parent(l_max, l, ni, C_ni, tree)
# update pi_new according to the found optimal pi_star
for item, node in pi_star.items():
pi_new[item] = node
d = min(d, l_max - l)
l = l + d
return pi_new
def assign_parent(l_max, l, ni, C_ni, tree):
"""implementation of line 5 of Algorithm 2
Returns:
updated \pi_{new}
Args:
l_max (int, required): the max level of the tree
l (int, required): current assign level
ni (node, required): a non-leaf node in level l-d
C_ni (item, required): item set whose ancestor is the non-leaf node ni
tree (tree, required): the old tree (\pi_{old})
"""
# get the children of ni in level l
children_of_ni_in_level_l = tree.get_children_given_ancestor_and_level(ni, l)
# get all the required weights
edge_weights = get_weights(C_ni, ni, children_of_ni_in_level_l, tree)
# assign each item to the level l node with the maximum weight
assign_dict = dict()
for ci, info in edge_weights.items():
assign_candidate_nodes = info[0]
assign_weights = np.array(info[1], dtype=np.float32)
sorted_idx = np.argsort(-assign_weights)
# assign item ci to the node with the largest weight
max_weight_node = assign_candidate_nodes[sorted_idx[0]]
if max_weight_node in assign_dict:
assign_dict[max_weight_node].append((ci, sorted_idx, assign_candidate_nodes, assign_weights))
else:
assign_dict[max_weight_node] = [(ci, sorted_idx, assign_candidate_nodes, assign_weights)]
edge_weights = None
# get each item's original assignment of level l in tree, used in rebalance process
origin_relation = dict()
for ci in C_ni:
origin_relation[ci] = tree.get_ancestor(ci, l)
# rebalance
max_assign_num = int(math.pow(2, l_max - l))
processed_set = set()
while True:
max_assign_cnt = 0
max_assign_node = None
for node in children_of_ni_in_level_l:
if node in processed_set:
continue
if node not in assign_dict:
continue
if len(assign_dict[node]) > max_assign_cnt:
max_assign_cnt = len(assign_dict[node])
max_assign_node = node
if max_assign_node == None or max_assign_cnt <= max_assign_num:
break
# rebalance
processed_set.add(max_assign_node)
elements = assign_dict[max_assign_node]
elements.sort(key=lambda x: (int(max_assign_node != origin_relation[x[0]]), -x[1]))
for e in elements[max_assign_num:]:
for idx in e[1]:
other_parent_node = e[2][idx]
if other_parent_node in processed_set:
continue
if other_parent_node not in assign_dict:
assign_dict[other_parent_node] = [(e[0], e[1], e[2], e[3])]
else:
assign_dict[other_parent_node].append((e[0], e[1], e[2], e[3]))
break
del elements[max_assign_num:]
pi_new = dict()
for parent_code, value in assign_dict.items():
max_assign_num = int(math.pow(2, l_max - l))
assert len(value) <= max_assign_num
for e in value:
assert e[0] not in pi_new
pi_new[e[0]] = parent_code
return pi_new
def get_weights(C_ni, ni, children_of_ni_in_level_l, tree):
"""use the user preference prediction model to calculate the required weights
Returns:
all weights
Args:
C_ni (item, required): item set whose ancestor is the non-leaf node ni
ni (node, required): a non-leaf node in level l-d
children_of_ni_in_level_l (list, required): the level l-th children of ni
tree (tree, required): the old tree (\pi_{old})
"""
edge_weights = dict()
for ck in C_ni:
edge_weights[ck] = list()
edge_weights[ck].append([]) # the first element is the list of nodes in level l
edge_weights[ck].append([]) # the second element is the list of corresponding weights
for node in children_of_ni_in_level_l:
path_to_ni = tree.get_parent_path(node, ni)
weight = 0.0
for n in path_to_ni:
sample_set = set() # the sample set that the target item is ck
# use the user preference prediction model to calculate the required weights.
# the detailed calculation process is omitted here
weight += calculate_weight_use_prediction_model(sample_set, n)
edge_weights[ck][0].append(node)
edge_weights[ck][1].append(weight)
return edge_weights