上篇文章记录了使用矩阵分解的方法来做个性化推荐。本篇文章参考了 http://book.paddlepaddle.org/08.recommender_system/ 上面的融合推荐算法，由于本人目前的精力放在学习MXNET上，所以将其用mxnet进行改写了。由于时间关系，暂时省略了原算法中对title的向量提取，预计下一篇会补充对title的CNN向量提取。

1. 数据准备

 MovieLens 百万数据集（ml-1m）。ml-1m 数据集包含了 6,000 位用户对 4,000 部电影的 1,000,000 条评价（评分范围 1~5 分，均为整数），由 GroupLens Research 实验室搜集整理。提示：原数据的分割符是'::', 可以用文本工具打开，替换成','或者'\t', 这样在使用pandas加载的时候能使用C库，否则加载数据非常慢。

1）movies.dat 每列分别是movieid, title, categories, 例如：1Toy Story (1995) Animation|Children's|Comedy。

需要注意的是电影类型也就是电影标签，一部电影可以有多个标签。比如上面的电影，类型是三种Animation，Children's，Comedy。

2）ratings.dat 每列分别是userid, movieid, score, time-stamp，表示用户对电影的评分。最后一列时间数据本例未使用

3）users.dat 是记录用户profile，分别是userid, gender, age 年龄段, job 职业类型，最后一列不知道是什么东东，本例未使用。

2. 自定义 DataIter和DataBatch，用于装载训练数据

跟上篇案例的类似，只是做了一点小变化，data.append(mx.nd.array(np.array(_data.tolist()),self.ctx)) import mxnet as mx import numpy as np class SimpleBatch(object): def __init__(self, data, label, pad=0): self.data = data self.label = label self.pad = pad class SimpleBatch(object): def __init__(self, data, label, pad=0): self.data = data self.label = label self.pad = pad class CustDataIter2: def __init__(self, data_names, data_gen, data_shape, label_names, label_gen, label_shape,ctx, batch_size,total_batches): self.data_names = data_names self.label_names = label_names self.data_gen = data_gen self.label_gen = label_gen self.data_shape = data_shape self.label_shape = label_shape self.batch_size = batch_size self.total_batches = total_batches self.cur_batch = 0 self.ctx = ctx def __iter__(self): return self def reset(self): self.cur_batch = 0 def __next__(self): return self.next() @property def provide_data(self): return zip(self.data_names,self.data_shape) @property def provide_label(self): return zip(self.label_names,self.label_shape) def next(self): if self.cur_batch < self.total_batches: data = [] for tdata in self.data_gen: _data = tdata[self.cur_batch*self.batch_size:(self.cur_batch+1)*self.batch_size] data.append(mx.nd.array(np.array(_data.tolist()),self.ctx)) assert len(data) > 0, "Empty batch data." label = [] for tlabel in self.label_gen: _label = tlabel[self.cur_batch*self.batch_size:(self.cur_batch+1)*self.batch_size] label.append(mx.nd.array(_label,self.ctx)) assert len(label) > 0, "Empty batch label." self.cur_batch += 1 # print('Return batch %d' % self.cur_batch) return SimpleBatch(data, label) else: raise StopIteration

3. 加载训练数据

分别有3个文件要装入，性别不是int型，所以需要进行变换。 def LoadRatingData(fname,delimiter='::'): df = pd.read_csv(fname,header=None,delimiter=delimiter,names=['userid','itemid','score','timestamp']) return df def gendercoding(x): if x =='M': return 2 else: return 1 def LoadUserData(fname,delimiter='::'): df = pd.read_csv(fname,header=None,delimiter=delimiter,names=['userid','gender','age','job','other'],index_col=0) df['gender'] = df['gender'].apply(gendercoding) return df def LoadItemData(fname,delimiter='::'): df = pd.read_csv(fname,header=None,delimiter=delimiter,names=['itemid','title','cat'],index_col=0) return df

4. encoding 电影类型

前面提到电影的类型是string，并且是多个值，例如Animation|Children's|Comedy。所以需要把这个值映射成一个向量来表示。 首先建立词典： def build_categories_dic(cate): cat_dic={} cat_index = {} cat_vocab = [] index = 0 for s in cate: s = s.split('|') for x in s: if cat_dic.get(x) == None: cat_dic[x]=1 cat_vocab.append(x) cat_index[x]=index index += 1 else: cat_dic[x] += 1 return (cat_dic,cat_vocab,cat_index) 然后把类型string，转换成对应的词典的index，再使用sklearn包的oneHotEncoder转成0，1编码的向量 def featureTags(catid): cat_dic,cat_vocab,cat_index = build_categories_dic(catid) encoded_cat=[] for cat in catid: s = cat.split('|') t = [] for tag in s: index = cat_index.get(tag) t.append(index) encoded_cat.append(t) df_cat = pd.DataFrame(encoded_cat).fillna(0) return df_cat def encodeTag(df): catid = df['cat'] itemid = df.index catfeatures = featureTags(catid).as_matrix() enc = preprocessing.OneHotEncoder(sparse=False) x_out = enc.fit_transform(catfeatures) s = [i for i in x_out] return pd.Series(s,index=itemid,name='encoded_cat')

5. 定义网络

参考原文的算法，暂时忽略了title这个属性 余弦相似度损失函数，并且rescale到1-5： def calc_cos_sim(a,b,min=1,max=5): x = mx.symbol.sum_axis( a*b, axis = 1) y = mx.symbol.sqrt(mx.symbol.sum_axis( mx.symbol.square(a), axis = 1)) * mx.symbol.sqrt(mx.symbol.sum_axis( mx.symbol.square(b), axis = 1)) cos = x/y rescale = cos*(max-min)+min return rescale 定义网络结构 def get_one_layer_mlp( max_userid, max_itemid,max_gender,max_age, max_job, k=64,): # user profile userid = mx.symbol.Variable('userid') gender = mx.symbol.Variable('gender') age = mx.symbol.Variable('age') job = mx.symbol.Variable('job') #times profile itemid = mx.symbol.Variable('itemid') # title = mx.symbol.Variable('title') cat = mx.symbol.Variable('cat') score = mx.symbol.Variable('score') # user latent features userid = mx.symbol.Embedding(data = userid, input_dim = max_userid, output_dim = k,name='userid_Embedding') userid = mx.symbol.FullyConnected(data = userid, num_hidden = k) userid = mx.symbol.Activation(data = userid, act_type="relu") gender = mx.symbol.Embedding(data = gender, input_dim = max_gender, output_dim = k/4,name='gender_Embedding') gender = mx.symbol.FullyConnected(data = gender, num_hidden = k/4) gender = mx.symbol.Activation(data = gender, act_type="relu") age = mx.symbol.Embedding(data = age, input_dim = max_age, output_dim = k/4,name='age_Embedding') age = mx.symbol.FullyConnected(data =age, num_hidden = k/4) age = mx.symbol.Activation(data = age, act_type="relu") job = mx.symbol.Embedding(data = job, input_dim = max_job, output_dim = k/2,name='job_Embedding') job = mx.symbol.FullyConnected(data =job, num_hidden = k/2) job = mx.symbol.Activation(data = job, act_type="relu") user = mx.symbol.concat(userid,gender,age,job,dim=1) # item latent features itemid = mx.symbol.Embedding(data = itemid, input_dim = max_itemid, output_dim = k,name='itemid_Embedding') itemid = mx.symbol.FullyConnected(data = itemid, num_hidden = k) itemid = mx.symbol.Activation(data = itemid, act_type="relu") cat = mx.symbol.FullyConnected(data = cat, num_hidden = k,name='cat_Fc') item = mx.symbol.concat(itemid,cat,dim=1) pred = calc_cos_sim(user,item,1,5) pred = mx.symbol.Flatten(data = pred) # loss layer pred = mx.symbol.LinearRegressionOutput(data = pred, label = score) return pred 网络结构如下：

6. 训练函数

def train(network,data_train_iter,data_valid_iter, context, num_epoch,learning, learning_rate): logging.getLogger().info('network debug:%s',network.debug_str()) model = mx.mod.Module( symbol = network, context=context, data_names=['userid', 'gender','age','job','itemid','cat'], label_names=['score'] ) model.fit(train_data = data_train_iter, eval_data =data_valid_iter, optimizer =learning, optimizer_params={'learning_rate':learning_rate}, eval_metric ='RMSE', num_epoch = num_epoch, ) return model def trainingModel(): TRAIN_DIR = 'C:/Users/chuanxie/PycharmProjects/mxnetlearn/data/movie/' ratingdf = LoadRatingData(TRAIN_DIR+'ml-1m/ratings.dat',delimiter='\t') userdf = LoadUserData(TRAIN_DIR+'ml-1m/users.dat',delimiter='\t') itemdf = LoadItemData(TRAIN_DIR+'ml-1m/movies.dat',delimiter='\t') np_encodedcat = encodeTag(itemdf) print ratingdf.shape,np_encodedcat.shape fulldf = ratingdf.join(userdf,on='userid').join(itemdf,on='itemid').join(np_encodedcat,on='itemid') '''reconstruct series to dataframe''' matrix_encoded_cat = fulldf['encoded_cat'].as_matrix() df_encoded_cat = np.array(matrix_encoded_cat.tolist()) print df_encoded_cat.shape data = np.array([fulldf['userid'],fulldf['gender'],fulldf['age'],fulldf['job'],fulldf['itemid'],fulldf['encoded_cat']]) print data.shape label = np.array([fulldf['score']]) context = mx.gpu() BATCH_SIZE = 16000 num_epoch = 100 trainIter = CustDataIter2(['userid', 'gender','age','job','itemid','cat'],data, [(BATCH_SIZE,),(BATCH_SIZE,),(BATCH_SIZE,),(BATCH_SIZE,),(BATCH_SIZE,),(BATCH_SIZE,df_encoded_cat.shape[1])], ['score'],label,[(BATCH_SIZE,)],context,BATCH_SIZE,data.shape[1]/BATCH_SIZE) max_userid = pd.Series(fulldf['userid']).max() max_itemid = pd.Series(fulldf['itemid']).max() max_gender = pd.Series(fulldf['gender']).max() max_age = pd.Series(fulldf['age']).max() max_job = pd.Series(fulldf['job']).max() net =get_one_layer_mlp( max_userid=max_userid, max_itemid=max_itemid,max_gender=max_gender, max_age = max_age , max_job = max_job, k=96) mx.viz.plot_network(net,shape={'userid':(128,),'gender':(128,),'age':(128,),'job':(128,),'itemid':(128,),'cat':(128,73)}) ##Train module train(net,trainIter,None,context,num_epoch=num_epoch,learning = 'adam',learning_rate=0.001) if __name__ == '__main__': trainingModel()

7. 运行结果

迭代100次就能把平均根方差降到0.7X，而前一篇文章单纯使用矩阵分解的算法迭代了1000次也只能达到0.9X，可见融合模型不仅加入了客户的行为数据，还融合了客户的自然属性，电影的自然属性，数据更加全面，因而准确度越高 INFO:root:Epoch[91] Train-RMSE=0.765856 INFO:root:Epoch[91] Time cost=7.528 INFO:root:Epoch[92] Train-RMSE=0.765861 INFO:root:Epoch[92] Time cost=7.575 INFO:root:Epoch[93] Train-RMSE=0.765705 INFO:root:Epoch[93] Time cost=7.794 INFO:root:Epoch[94] Train-RMSE=0.765217 INFO:root:Epoch[94] Time cost=7.455 INFO:root:Epoch[95] Train-RMSE=0.765055 INFO:root:Epoch[95] Time cost=7.401 INFO:root:Epoch[96] Train-RMSE=0.765331 INFO:root:Epoch[96] Time cost=7.409 INFO:root:Epoch[97] Train-RMSE=0.765364 INFO:root:Epoch[97] Time cost=7.401 INFO:root:Epoch[98] Train-RMSE=0.765842 INFO:root:Epoch[98] Time cost=7.425 INFO:root:Epoch[99] Train-RMSE=0.766367 INFO:root:Epoch[99] Time cost=7.445