With the explosive growth of information available online, Recommender System (RS
), as a useful information filtering tool, is used for estimating users’ preferences on items they have not seen and guides them to discover products or services they might be interested in from massive possible options. In general, recommender systems are classified into the following three categories based on the forms of recommendations[Balabanović and Shoham1997]: Collaborative recommendations, Content-based recommendations and Hybrid recommendations. Collaborative recommendations make users recommended items that people with similar tastes preferred in the past. Content-based recommendations make users recommended items similar to the ones the user preferred in the past. Hybrid recommendations integrates two or more types of recommendation strategies, which helps to avoid certain limitations of single strategy [Adomavicius and Tuzhilin2005].
Recommender System increasingly plays an essential role in industry area, which promotes services for many applications [Gomez-Uribe and Hunt2016, Davidson and Liebald2010]. In addition, in order to help customers find exactly what they need, recommendation techniques have been studied and deployed extensively on E-commerce platforms, which provides good user experience and promotes incredible increment in revenue. Usually, the deployed framework for our online E-commerce platforms is illustrated in Fig. 1. Specifically, when a user visits it through a terminal, such as smart phones, the system firstly analyzes his/her long and short term behaviors and then his/her interested items, called Triggers, are selected. Then, massive items closely related with Triggers are generated. Further, top K(e.g., 500) of them (based on the “matching score”), along with extra information (e.g., user, item, user-item cross features), are delivered to the next Ranking stage where it mainly contains two core modules, namely CTR and CVR. Finally, the recommendation results are generated and displayed to the user. In this paper, we mainly focused on the CVR module.
In the past few decades, deep learning has been witnessed the tremendous successes in many application areas[Zhang, Yao, and Sun2017, Liu et al.2017], such as image classification [Krizhevsky, Sutskever, and Hinton2012, Simonyan and Zisserman2014] , speech recognition [Deng et al.2010, Deng et al.2013] and object detection [Girshick2015, Ren et al.2015]. Meanwhile, recent studies also demonstrate its efficiency and effectiveness in coping with recommendation tasks[Elkahky, Song, and He2015, Wang, Wang, and Yeung2015, Chen et al.2017, He et al.2017, Huang et al.2015, Yang et al.2017, Guo et al.2017]. Though deep learning has been partially overcoming obstacles of conventional models and gaining momentum due to its state-of-the-art performances, it has apparent deficiencies, such as a huge amount of data and powerful computational facilities required for training, more importantly many hyper-parameters to be tuned. Recently, gcForest [Zhou and Feng2017], an alternative to DNN, is proposed, which generates a deep forest ensemble, with a cascade structure to do representation learning. In addition, it achieves highly competitive performance compared with DNN for various domains’ tasks while having fewer hyper-parameters.
In this paper, partially inspired by gcForest [Zhou and Feng2017], we firstly propose a multi-Level Deep Cascade Trees model (short as ldcTree) to cope with the essential task CVR prediction in the “also view” module. ldcTree is another alternative to DNN
and encourages to do representation learning by a level-by-level cascade structure. Specifically, it takes a multi-dimensional representational feature vector from preceding level and outputs its processing results to the next level by employing Gradient Boosting Decision Trees (GBDT ) models [Friedman2001], where a new tree is created to model the residual of previous trees during each iteration, and a traversal from root node to a leaf node represents a combination rule of certain input features. One step further, we propose to utilize the cross-entropy value of each leaf node on the trees in the preceding level GBDT to construct the feature representation for the next level GBDT, which results a clear explanation of the spliting node in the next level GBDT, i.e., a traversal from root to a leaf node in next level GBDT indicates a combination rule of certain paths on the trees from preceding level GBDT. Then, inspired by the idea of ensemble learning, we proposed Ensemble ldcTree (short as E-ldcTree), which encourages the model’s diversity and enhances the representation ability.
Furthermore, it is noteworthy that a small number of raw features contributes the majority of explanatory power while the remaining features have only a marginal contribution in GBDT models [He et al.2014], which results in the importance of certain raw features can’t be demonstrated. Therefore, we further proposed an improved Feature learning method based on the above EldcTree, named F-EldcTree, which takes more adequately use of weak and strong correlation features identified by pre-trained GBDT model at corresponding levels. The key contributions of this paper are:
We propose a novel model ldcTree and its extension , which are decision tree ensemble methods by exploiting the deep cascade structures and using a cross-entropy based feature representation. It exhibits a strong feature representation ability and has a clear explanation.
To the extent of our knowledge, our proposed F-EldcTree is the first recommendation work which adequately takes into account weak and strong correlation features identified by pre-trained GBDT model at corresponding levels, and contributes more excellent ability for representation learning.
We have successfully deployed the proposed methods to the recommendation module in our E-commerce platform, and carry out online experiments with more than 100M users and items, furthermore achieves 12 percent CVR improvement compared with the baseline model.
The rest of this paper is organized as follows. Section 2 briefly reviews existing related work. Section 3 describes the proposed approach in detail, followed by presenting experimental results on both off-line evaluation and online applications in Section 4. We conclude the paper in Section 5.
Conversion Rate Prediction
Conversions are very rare events and only a very small portion of the users will eventually convert after clicking or being shown, resulting in extremely challenging for building thus models [Mahdian and Tomak2007, Chapelle, Manavoglu, and Rosales2015, Rosales, Cheng, and Manavoglu2012, Oentaryo et al.2014]. In addition, it can be broadly divided to two categories of Post View Conversion (PVC) and Post Click Conversion (PCC), which means conversion after viewing an item without having clicked it itself and conversion after having clicked, respectively. In the context of this paper, conversion refers to the purchase event that occurs after a user clicking an item, i.e. post click conversions (PCC) [Rosales, Cheng, and Manavoglu2012].
Tree based Feature Representation
GBDT follows the Gradient Boosting Machine (GBM) [Friedman2001], which produces competitive, highly robust, interpretable procedures for both regression and classification, especially appropriate for mining less than clean data. In literature [He et al.2014], a hybrid model, combining GBDT
with Logistic Regression (LR), outperforms either of these methods on their own. It treats each individual tree as a bin feature and takes the index of the leaf node an instance ends up falling in as value. Therefore, it converts a real-valued vector into a compact binary-valued vector, further included into the next linear model, i.e. LR. Compared with [He et al.2014], our proposed method employs the cross-entropy based feature representation in a deep cascaded structure, which results in strong and explainable representation ability, i.e., a traversal from root to a leaf node in next level GBDT indicates a combination rule of certain paths on the trees from preceding level GBDT.
, an alternative to deep neural networks for many tasks, is proposed in literature[Zhou and Feng2017], which employs deep forest structure to do representation learning. Specifically, it takes a multi-dimensional class vector from preceding level, together with the original feature vector, as the inputs of next level. Our proposed methods mainly have two significant differences from gcForest:
, the class-specific prediction probabilities form a feature vector, further included into the next level forest after concatenating it with the original features. However, We employGBDT as the base unit in the proposed deep cascade structure. In addition, we use the cross-entropy in each tree of the preceding GBDT as the feature representation for next level GBDT. The aforementioned two points lead to a more explainable feature representation ability of the proposed method, e.g., a traversal from root to a leaf node in next level GBDT indicates a combination rule of certain paths on the trees from preceding level GBDT.
Compared with gcForest, Our method takes into account mutual complementarity between strong correlation features and weak correlation features for better representation learning.
The Proposed Approach
In this section, we firstly proposed a novel multi-Level Deep Cascade Trees (ldcTree) to tackle the CVR prediction problem in recommendation. Specifically, the base structure of the ldcTree is constructed by stacking several GBDTs sequentially, and the cross-entropy of each leaf node in the preceding GBDT is calculated and used to form the input feature representation for the next GBDT. Moreover, inspired by the idea of ensemble learning, an improved structure of the ldcTree is proposed, named Ensemble ldcTree (EldcTree), which further encourages the model’s diversity and enhances the representation ability. Finally, we notice that a small number of features contributes the majority of explanatory power while the remaining features have only a marginal contribution in GBDT models [He et al.2014], which leads to the importance of weak correlation features, especially the combination of weak correlation features, can’t be revealed. Therefore, based on EldcTree, We propose an improved Feature learning method, named F-EldcTree. We will present the details in the following parts.
Representation Learning based on Cascade Trees
Inspired by representation learning in deep neural network which mostly relies on the level-by-level abstraction of features, we propose a novel method named ldcTree by employing the deep cascade tree structure. In ldcTree, level-by-level greedily learning towards the final target is carried out. Specifically, each level with a certain number of trees included in GBDT, receives the outputted features from its proceding level. Referring to Fig. 2, it is an illustration of ldcTree, where it contains two levels, with each level three trees and two trees respectively. To facilitate the narration, some mathematical notations are defined as follows:
: the cross-entropy of the -th leaf node of the -th tree at level .
: the number of instances falling in the -th leaf node of the -th tree at level .
: the split threshold for the k-th node of the j-th tree at the level .
: the number of leaf nodes of the -th tree at level .
: the feature value at the j-th dimension of the feature at level .
: the number of trees for the GBDT model at level .
: the predicted probability of the n-th instance on the j-th tree at the level .
: the ground truth label of the n-th instance : 0 or 1 in our two class CVR problem.
Given an instance, according to the principle of GBDT model, each individual tree will produce a path from the root node to a leaf node. Instances are split into different paths and each leaf node will gather a certain part of them. Then, we define the cross-entropy at each leaf node as:
Therefore, there are cross-entropy values on the -th tree on the -th level, i.e., and each of them is a possible instantiation of . For all the trees in a GBDT at the level , we denote the feature representation as , and use it as the inputs of the GBDT at the level .
Without loss of generality, assuming the feature with the best gini value is chosen for splitting instances on the -th node of the -th tree at the level . The split threshold divides the set of , into two subsets :
where is an indicator faction that outputs 1 if is true and zero otherwise.
By using cross-entropy as the basic feature representation for leaf nodes, the proposed has a clear explanation: i.e., a traversal from root to leaf nodes in the next level GBDT corresponds to the combination of certain traversals in the preceding GBDT, which also leads to a distributed feature representation ability.
Explanation: Usually, in Eq.(2) and Eq.(3) takes a value from the set of . The splitting process is carried out repeatedly until the stop rule for leaf nodes holds (We show an example in Fig. 2). It can be seen that, each element in or corresponds a path on the -th tree of level . Therefore, each splitting node at the level corresponds a split of paths in a certain tree at the preceding level . Consequently, a path at the level corresponds a union of several paths at the preceding level . It indicates the nature of the proposed : a clear explanation and a distributed feature representation ability. As shown in Fig. 2, an instance is represented as a three-dimensional feature vector after level 1 with three trees, where takes a value from the set of 0.13, 0.18, 0.19, 0.15, analogously for and . Given as the inputs at level , is chosen as the split feature with the split threshold on Tree 1, which leads to and . And analogously, and . Consequently, the union of paths at leaf node A can be represented as , i.e. , and analogously for Tree 2.
Moreover, inspired by the ensemble learning idea, we propose an Ensemble ldcTree named EldcTree by constructing several parallel ldcTrees to enhance the representation ability further. As shown in Fig. 3, each horizontal part of the ensemble structure is a single ldcTree, and each vertical part of the ensemble structure are parallel GBDTs. Their initial input features for the first level of each ldcTree are different and chosen randomly from a raw feature pool which encourages the model’s diversity. The last level’s outputted features from each ldcTree are concatenated together and used as the input features of the last GBDT for final prediction. Here, the diversity in our model not only represents the nature of diverse abstracted high-level features, but also the value of inherent ensemble learning. For example, each individual Horizontal structure randomly choose a subset of features from the feature pool as input, which are then abstracted to different high-level features through the cascaded structure. It inherits the idea in feature engineering by learning high-level features from different combinations of low-level raw features. Then, the final GBDT model, concatenating all the high-level features from preceding ldcTree as inputs, indeed follows the ensemble learning idea for further enhancing the performance of the entire model. It is noteworthy that such an ensemble structure is naturally apt to parallel implementation and have the potential for incremental learning, e.g., the idea used in broad learning system [Chen and Liu2018]. We leave it as a future work.
Cascade Trees Associated With Weak Correlation Features
Though features can be chosen randomly from a given feature pool and used as the input features of each ldcTree, features indeed have different importance for prediction. In this paper, we use the statistic Boosting Feature Importance [He et al.2014], which aims to capture the cumulative loss reduction attributable to a feature, to measure feature’s importance. More specifically, a best feature is selected and split to maximize the squared error reduction during each tree node construction. Therefore, the importance of each feature is determined by summing the total reduction for itself across all the trees. Typically, a small number of features contributes the majority of explanatory power while the remaining features have only a marginal contribution in GBDT models. Here, we regard the features contributing the majority of explanatory power as SCF, i.e. “Strong Correlation Features”, while features having only a marginal contribution as WCF, i.e. “Weak Correlation Features”.
To take adequate use of WCF, we proposed an improved Feature learning method based on the above EldcTree, named F-EldcTree, whose structure is showed in Fig. 4. In Fig. 4(a), a separated GBDT is pre-trained to identify the importance of all the initial raw features, and these features are further split into two subsets, namely WCF and SCF. Due to our practical lessons, single feature from WCF contributes little to final prediction results, while the combinations of these features from WCF not. Moreover, in GBDT models, a traversal from root to a leaf node in each tree indicates a combination rule of certain raw features. Therefore, we can resort to GBDT models for uniting certain features from WCF to take full advantage of them and collaborate with SCFs to further improve the prediction accuracy.
The strategies of randomly selecting features in gcForest [Zhou and Feng2017] convinces us its effectiveness in ensemble learning. Therefore, in Fig. 4(b), features are also randomly chosen from WCFs during the first level GBDT training stage of each ldcTree, which not only can learn the combination of certain WCFs, but also encourages the model’s diversity and enhances the representation ability further. For GBDTs at the remaining levels of each ldcTree, their input features consists of two parts: one is the representational features from the preceding level, the other is randomly chosen from the SCFs. In this way, the F-EldcTree starts learning features from WCFs and combines the learned features (i.e., the combinations of WCFs) with SCFs little by little. Lastly, an additional GBDT model concatenates all the representational features from ldcTrees as the inputs for final results.
To evaluate the effectiveness of the proposed method, we conducted extensive experiments including off-line and online evaluations. First, we present the evaluation settings including data set preparation, evaluation metrics, a brief description of related comparison methods, and the hyper-parameters settings of the proposed method. Then, we present the experimental results of different methods on the off-line data set along with the analysis. Finally, we present experimental results of different methods for online deployment through A/B test.
Data set preparation
The off-line benchmark data set was constructed from the real click and purchase logs of our recommendation module in several consecutive days of December, 2017. And it consists of more than 100M instances, each of which contains user/item features and label (here, an individual id for each instance, while label is positive if purchase after clicking, or negative if no purchase after clicking). When preparing off-line benchmark data set, we divided the whole benchmark data into three disjoint parts according to the id, i.e., 40, 20 and 40 percent of the whole benchmark data for training data, validation data and test data respectively. Additionally, we extract hundreds of raw features including user features, item features and user-item cross features for each instance. For example, user features include users’ ages, genders and purchasing powers, etc. Item features include items’ prices, historical CVRs and Click-through Rate (CTRs), etc. User-item cross features include uses’ historical CVRs, preference scores, etc, on items.
In order to comprehensively compare the performance of the propose method with other related methods, we adopt two commonly used metrics namely AUC (Area Under Curve) and
score based on precision and recall. Specifically, denoting all ground truth positive instances asT and all predicted positive instances as P, then precision and recall are defined as follows:
Then, the score is defined as:
Related comparison methods
In the following experiments, we compare the proposed method with other related methods including:
Naive GBDT: We refer to a single GBDT model without level-by-level learning as Naive GBDT.
GBDT + LR: First, feature representation is trained from a GBDT model. Then, it is used for CVR prediction by a Logistic Regression (LR). For feature representation, it calculated a bin feature for each individual tree by taking the index of the leaf node which an instance ends up falling in as the corresponding feature value. Therefore, it converts a real-valued raw feature vector into a compact binary-valued feature vector [He et al.2014].
DNN: Referring to [He et al.2017], we design a DNN
structure including three hidden layers and a prediction layer, where ReLU is used as the activation function for each hidden layer. We choose the hyper-parameters on the validation set. For instance, the number of units for each hidden layer is set as 128. Dropout rate is set as 0.5. We use the cross-entropy loss andSGD algorithm to train this DNN model.
gcForest: Following our practical experience, we replace the Forests in the original gcForest [Zhou and Feng2017] with GBDTs. Since the CVR prediction is a binary classification problem, a two-dimensional class-specific feature vector is obtained from each GBDT. Then, it is used along with the raw feature vector as the inputs of next level GBDT
for learning deep feature representation further.
Moreover, we also evaluate the performance among our proposed ldcTree, EldcTree and F-EldcTree methods for demonstrating the effectiveness of ensemble learning and the adequate use of WCF and SCF at corresponding levels. It should also be noted that all methods use the same raw features as inputs if not specified (e.g., here, user, item and user-item cross features are included.)
the type of loss function
|minimum instance numbers when node split||20|
|sampling rate of train set for each iteration||0.6|
|sampling rate of features for each iteration||0.6|
|the tree depth||8|
|the number of trees||150|
We choose the hyper-parameters of the proposed methods according to the AUC metric on the validation set. And the main hyper-parameters of the proposed methods we used in all the following concern experiments are shown in Tab. 1. Here, we take a critical parameter “the tree depth” as example to illustrate the process of parameter selection in ldcTree model. After sampling a small subset from the whole data, we train different ldcTree models by changing the tree depth while fixing other settings. Results are shown in Tab. 2. It can be seen that the time cost for each iteration increases consistently with the growth of the tree depth, and the corresponding AUC on the validation set also increases. In addition, the time cost increases significantly when the depth increases from 8 to 10, while for the AUC value, it grows at a snail’s pace. For example, marginal gains of 0.003 are achieved by increasing the depth from 8 to 10. Therefore, to achieve a trade-off between model capacity and complexity, we set the hyper-parameter “the tree depth” as 8. Moreover, EldcTree and F-EldcTree can refer to the same hyper-parameters as ldcTree. Finally, aforementioned naive experiments also taught us that the performance of models can’t be boosted further while increasing the tree depth blindly.
|Metric Name||Depth 4||Depth 6||Depth 8||Depth 10|
Comparison results on off-line dataset
We report the AUC values and scores of different methods on the off-line test set. The results are shown in Tab. 3 and Tab. 4, respectively. Referring to Tab. 3, it can be seen that the GBDT + LR method achieves a gain of 0.0039 compared to Naive GBDT due to the additional LR classifier. DNN and gcForest achieve better results than GBDT + LR. It convinces us that strong representation features are learned due to their deep structures. Our proposed method ldcTree achieves higher AUC values than certain related methods(such as Naive GBDT, GBDT+LR, DNN). Moreover, the proposed EldcTree achieves better AUC result than ldcTree due to the idea of ensemble learning. Finally, the further proposed F-EldcTree achieves the best result than all the other competitive methods by taking full use of WCF and SCF little by little, together with the idea of ensemble learning. And the gain is nearly 0.063 compared to the initial baseline Naive GBDT. According to our practical lessons, it should be noted that a gain of 0.01 in off-line AUC can lead to big increment in revenue in our online recommendation system. In conclusion, the significant gain over initial Naive GBDT convinces the effectiveness of the proposed deep cascade structure for stronger feature representation, and the gain over gcForest convinces the effectiveness of level-by-level learning, for example, taking the outputs of preceding level as the inputs of the next level. Moreover, compared with EldcTree, results of F-EldcTree convince the idea by taking full use of weak and strong correlation features.
As for the score, we report several values by setting different thresholds. First, we sort all the instances in a descending order according to the predicted score. Then, we choose 3 thresholds namely top@10%, top@20% and top@50% to split the predictions into positive and negative groups accordingly. Finally, we calculate the Precision, Recall and scores of these predictions at different thresholds. Results are showed in Tab. 4. As can be seen, the proposed method achieves the best performance which is consistent with Tab. 3.
|GBDT + LR||0.7731|
|Method Type||Method Name||top@10 percent||top@20 percent||top@50 percent|
|Compare Methods||Naive GBDT||5.75%||36.39%||9.93%||4.16%||51.69%||7.70%||2.42%||76.57%||4.68%|
|GBDT + LR||6.39%||37.40%||10.92%||4.63%||54.16%||8.52%||2.68%||78.71%||5.19%|
Online evaluation results
|Model Name||Day 1||Day 2||Day 3||Day 4|
Next, we firstly present the effectiveness of level-by-level learning through online contrastive experiments. Then, we further demonstrate the effectiveness of F-EldcTree by taking full advantage of WCF and SCF compared with other competitors. It should also be noted that all comparison methods use the same input features if not specified (e.g., here, user, item and user-item cross features are included.). In addition, we fix all the other online recommendation modules unchanged except the CVR module.
The Effectiveness of level-by-level Learning
For demonstrating the effectiveness of the proposed level-by-level learning, we implement ldcTree and EldcTree methods with the same features from Naive GBDT. After deploying them in the online recommendation system, we record four days’ purchase logs and calculate the relative increasement in CVR. The A/B test results are showed in Tab. 5. It can be seen the proposed ldcTree method achieves more than 4% gain of CVR averagely, while EldcTree more than 7%. In addition, after analyzing the difference between the structures of the two methods, we find that the gain mainly comes from the stronger feature representation ability of the proposed deep cascade structure in EldcTree. It is consistent with the experimental results on the aforementioned off-line data set.
The Effectiveness of F-EldcTree.
After demonstrating the effectiveness of utilizing level-by-level learning through online A/B experiments, we employed F-EldcTree for online environment. In addition, the features for other competitive methods are exactly the same with F-EldcTree. The A/B test results are showed in Fig. 5, where gcForest and DNN achieve better results than Naive GBDT due to their deep feature representation abilities. As for the proposed method, it achieves the best result, i.e. 12 percent increment in CVR among all the methods.
In a nutshell, considering the experimental results from both off-line and online tests, we conclude that the proposed method has a stronger feature representation ability due to its deep cascade structure and the adequate use of WCFs and SCFs at corresponding levels. Moreover, these two distinct characteristics enables the learned features to have a clear explanation as depicted in Section 3.1.
Conclusions and future work
In this paper, we introduce effective and efficient distributed feature learning methods ldcTree and its extension EldcTree, which have a deep cascade structure by stacking several GBDT units sequentially. By using a cross-entropy based feature representation, it leads to a clear explanation and a distributed feature representation ability. Moreover, after taking into account mutual complementarity between strong correlation features and weak correlation features under the ensemble learning framework, the proposed method F-EldcTree achieves the best performance in both off-line and on-line experiments. Specifically, we successfully deploy the proposed method online in our E-commerce platform, it achieves a significant improvement compared with the previously baseline, i.e. 12 percent increment in CVR. Our methods have small training cost and are naturally apt to parallel implementation. In addition, it is promising to be applicable for other online advertising scenarios.
Future work may include the following two directions: 1) Incorporating more features, such as information from the parent nodes and sibling nodes for learning stronger feature representation. 2) Studying the end-to-end training method for jointly feature learning and classifying based on the proposed deep cascade tree structure.
This work was partly supported by the National Natural Science Foundation of China (NSFC) under Grants 61806062 and 61751304.
- [Adomavicius and Tuzhilin2005] Adomavicius, G., and Tuzhilin, A. 2005. Toward the next generation of recommender systems: A survey of the state-of-the-art and possible extensions. IEEE Transactions on Knowledge & Data Engineering (6):734–749.
- [Balabanović and Shoham1997] Balabanović, M., and Shoham, Y. 1997. Fab: content-based, collaborative recommendation. Communications of the ACM 40(3):66–72.
- [Chapelle, Manavoglu, and Rosales2015] Chapelle, O.; Manavoglu, E.; and Rosales, R. 2015. Simple and scalable response prediction for display advertising. ACM Transactions on Intelligent Systems and Technology (TIST) 5(4):61.
- [Chen and Liu2018] Chen, C. P., and Liu, Z. 2018. Broad learning system: an effective and efficient incremental learning system without the need for deep architecture. IEEE transactions on neural networks and learning systems 29(1):10–24.
- [Chen et al.2017] Chen, J.; Zhang, H.; He, X.; Nie, L.; Liu, W.; and Chua, T.-S. 2017. Attentive collaborative filtering: Multimedia recommendation with item-and component-level attention. In Proceedings of the 40th International ACM SIGIR conference on Research and Development in Information Retrieval, 335–344. ACM.
- [Davidson and Liebald2010] Davidson, J., and Liebald, B. 2010. The youtube video recommendation system. In Proceedings of the fourth ACM conference on Recommender systems, 293–296. ACM.
- [Deng et al.2010] Deng, L.; Seltzer, M. L.; Yu, D.; Acero, A.; Mohamed, A.-r.; and Hinton, G. 2010. Binary coding of speech spectrograms using a deep auto-encoder. In Eleventh Annual Conference of the International Speech Communication Association.
- [Deng et al.2013] Deng, L.; Li, J.; Huang, J.-T.; Yao, K.; Yu, D.; Seide, F.; Seltzer, M.; Zweig, G.; He, X.; Williams, J.; et al. 2013. Recent advances in deep learning for speech research at microsoft. In Acoustics, Speech and Signal Processing (ICASSP), 2013 IEEE International Conference on, 8604–8608. IEEE.
- [Elkahky, Song, and He2015] Elkahky, A. M.; Song, Y.; and He, X. 2015. A multi-view deep learning approach for cross domain user modeling in recommendation systems. In Proceedings of the 24th International Conference on World Wide Web, 278–288. International World Wide Web Conferences Steering Committee.
- [Friedman2001] Friedman, J. H. 2001. Greedy function approximation: a gradient boosting machine. Annals of statistics 1189–1232.
Proceedings of the IEEE international conference on computer vision, 1440–1448.
- [Gomez-Uribe and Hunt2016] Gomez-Uribe, C. A., and Hunt, N. 2016. The netflix recommender system: Algorithms, business value, and innovation. ACM Transactions on Management Information Systems (TMIS) 6(4):13.
- [Guo et al.2017] Guo, H.; Tang, R.; Ye, Y.; Li, Z.; and He, X. 2017. Deepfm: A factorization-machine based neural network for ctr prediction. arXiv preprint arXiv:1703.04247.
- [He et al.2014] He, X.; Pan, J.; Jin, O.; Xu, T.; Liu, B.; Xu, T.; Shi, Y.; Atallah, A.; Herbrich, R.; Bowers, S.; et al. 2014. Practical lessons from predicting clicks on ads at facebook. In Proceedings of the Eighth International Workshop on Data Mining for Online Advertising, 1–9. ACM.
- [He et al.2017] He, X.; Liao, L.; Zhang, H.; Nie, L.; Hu, X.; and Chua, T.-S. 2017. Neural collaborative filtering. In Proceedings of the 26th International Conference on World Wide Web, 173–182. International World Wide Web Conferences Steering Committee.
- [Huang et al.2015] Huang, W.; Wu, Z.; Chen, L.; Mitra, P.; and Giles, C. L. 2015. A neural probabilistic model for context based citation recommendation. In AAAI, 2404–2410.
- [Krizhevsky, Sutskever, and Hinton2012] Krizhevsky, A.; Sutskever, I.; and Hinton, G. E. 2012. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, 1097–1105.
- [Liu et al.2017] Liu, W.; Wang, Z.; Liu, X.; Zeng, N.; Liu, Y.; and Alsaadi, F. E. 2017. A survey of deep neural network architectures and their applications. Neurocomputing 234:11–26.
- [Mahdian and Tomak2007] Mahdian, M., and Tomak, K. 2007. Pay-per-action model for online advertising. In Proceedings of the 1st international workshop on Data mining and audience intelligence for advertising, 1–6. ACM.
- [Oentaryo et al.2014] Oentaryo, R. J.; Lim, E.-P.; Low, J.-W.; Lo, D.; and Finegold, M. 2014. Predicting response in mobile advertising with hierarchical importance-aware factorization machine. In Proceedings of the 7th ACM international conference on Web search and data mining, 123–132. ACM.
- [Ren et al.2015] Ren, S.; He, K.; Girshick, R.; and Sun, J. 2015. Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances in neural information processing systems, 91–99.
- [Rosales, Cheng, and Manavoglu2012] Rosales, R.; Cheng, H.; and Manavoglu, E. 2012. Post-click conversion modeling and analysis for non-guaranteed delivery display advertising. In Proceedings of the fifth ACM international conference on Web search and data mining, 293–302. ACM.
- [Simonyan and Zisserman2014] Simonyan, K., and Zisserman, A. 2014. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556.
- [Wang, Wang, and Yeung2015] Wang, H.; Wang, N.; and Yeung, D.-Y. 2015. Collaborative deep learning for recommender systems. In Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 1235–1244. ACM.
[Yang et al.2017]
Yang, C.; Bai, L.; Zhang, C.; Yuan, Q.; and Han, J.
Bridging collaborative filtering and semi-supervised learning: a neural approach for poi recommendation.In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 1245–1254. ACM.
- [Zhang, Yao, and Sun2017] Zhang, S.; Yao, L.; and Sun, A. 2017. Deep learning based recommender system: A survey and new perspectives. arXiv preprint arXiv:1707.07435.
[Zhou and Feng2017]
Zhou, Z.-H., and Feng, J.
Deep forest: towards an alternative to deep neural networks.
Proceedings of the 26th International Joint Conference on Artificial Intelligence, 3553–3559. AAAI Press.