Parallel matrix factorization for recommender systems

Matrix factorization, when the matrix has missing values, has become one of the leading techniques for recommender systems. To handle web-scale datasets with millions of users and billions of ratings, scalability becomes an important issue. Alternating least squares (ALS) and stochastic gradient descent (SGD) are two popular approaches to compute matrix factorization, and there has been a recent flurry of activity to parallelize these algorithms. However, due to the cubic time complexity in the target rank, ALS is not scalable to large-scale datasets. On the other hand, SGD conducts efficient updates but usually suffers from slow convergence that is sensitive to the parameters. Coordinate descent, a classical optimization approach, has been used for many other large-scale problems, but its application to matrix factorization for recommender systems has not been thoroughly explored. In this paper, we show that coordinate descent-based methods have a more efficient update rule compared to ALS and have faster and more stable convergence than SGD. We study different update sequences and propose the CCD++ algorithm, which updates rank-one factors one by one. In addition, CCD++ can be easily parallelized on both multi-core and distributed systems. We empirically show that CCD++ is much faster than ALS and SGD in both settings. As an example, with a synthetic dataset containing 14.6 billion ratings, on a distributed memory cluster with 64 processors, to deliver the desired test RMSE, CCD++ is 49 times faster than SGD and 20 times faster than ALS. When the number of processors is increased to 256, CCD++ takes only 16 s and is still 40 times faster than SGD and 20 times faster than ALS.     

Shrinkage and warpage prediction of injection‐molded thin‐wall parts using artificial neural networks

This study demonstrates the successful use of back‐propagation artificial neural networks (BPANNs) in predicting the shrinkage and warpage of injection‐molded thin‐wall parts. The effects of structural parameters of a BPANN on the predictionaccuracy and the capability of a BPANN in determining the optimal process condition are also discussed. The training and testing data are obtained experimentally based on a Taguchi L₂₇ (3¹³) test schedule. The results show that the trained BPANN can successfully predict the shrinkage and warpage of injection‐molded thin‐wall parts. Comparing the prediction accuracies of the trained BPANN and C‐Mold software, it is noted that the trained BPANN predicts more accurately. In terms of determining the optimal process condition for minimizing the shrinkage and warpage of injected thin‐wall parts, the trained BPANN is also shown to give a better optimal process condition than Taguchi's method. Polym. Eng. Sci. 44:2029–2040, 2004. © 2004 Society of Plastics Engineers.     

Optimal process conditions of shrinkage and warpage of thin‐wall parts

Optimal process conditions of thin‐wall injection molding of a cellular phone cover were investigated with the consideration of interaction effects between process parameters. L₂₇ experimental tests based on Taguchi's method were performed, and then Cyclone Scanner, PolyCAD and PolyWorks were used to measure the shrinkage and warpage of the thin‐wall injected parts to determine the optimal process conditions. Based on the results of the analysis of variables and the F‐test, interaction effects for each observed factor were determined. The results indicated that the packing pressure was the most important process parameter affecting the shrinkage and warpage of the thin‐wall part. The optimal process conditions were different for the shrinkage and the warpage. This was because during the injection process, the mechanisms affecting shrinkage or warpage were different. Compared with the results obtained with simplified thin‐wall parts in the literature, it was found that the geometry of a real commercial part did affect the optimal process conditions and the order of influence of process parameters. The optimal process conditions determined by Taguchi's method for reducing the shrinkage and warpage were verified experimentally in this work. Polym. Eng. Sci. 44:917–928, 2004. © 2004 Society of Plastics Engineers.     

Influence of Potassium on Preparation and Performance of PTC Resistors

Different amounts of K₂CO₃ were added to (Ba,Sr)TiO₃-based PTCR (positive temperature coefficient of resistance) ceramics to investigate their influence on the microstructural and electrical properties. Experimental results showed that the incorporation of K acted as an A-site acceptor-type dopant. In addition to enhancing discontinuous grain growth, the increase of K₂CO₃ was found to raise the room-temperature resistivity which was dominated by grain-boundary resistance rather than grain resistance. By adjusting to a suitable amount of donor dopant, the inherent contamination of K in raw material can be compensated to achieve a high-quality PTC resistor.     

Solute diffusion through degradable semicrystalline polyethylene glycol/poly(l-lactide) copolymers

Summary The diffusion of C.I. direct orange 34(MW=299) and benzoic acid(MW=122) through degraded semicrystalline polyethylene glycol(PEG)/poly(L-lactide)(PLLA) block copolymers with various PEG contents and PEG segment lengths at 37°C was studied by UV-visible spectroscopy, differential scanning calorimetry(DSC), wide angle X-ray diffractometer (WAXS) and scanning electron microscopy(SEM). The influences of the PEG contents, PEG segment lengths and hydrolytic degradation of PEG/PLLA copolymers on the solute diffusion coefficient and mode for transport were investigated. It is concluded that the diffusion rate increases with the increase of PEG contents and PEG segment lengths in PEG/PLLA copolymers. This is understandable that the increase of PEG content and PEG segment length both make the degree of crystallinity decrease. The steady state of mass flux could not be reached at the diffusion times up to 1000 h, because the copolymers underwent hydrolysis reaction during this period. Furthermore, it is understood that the characteristic time of diffusion as defined by the square of film thickness at an instant of time over the diffusion coefficient of solute through polymer decreases with the increasing diffusion time.     

Thermally reversible cross-linked polyamides and thermo-responsive gels by means of Diels-Alder reaction

Cross-linked polyamides and polyamide gels were prepared from maleimide-containing polyamides and a tri-functional furan compound and showed thermal reversibility in cross-linking behavior and in gel formation through Diels-Alder (DA) and retro-DA reactions. The rate constant k of the DA cross-linking reaction were 1.25-4.83×10⁻⁵ dm³ mol⁻¹ s⁻¹ in the temperature range of 20-60 °C with an activation energy of 32.1 kJ mol⁻¹. The cross-linking densities, thermal properties, and thermal reversibility of the polyamides/furan polymers were adjustable with the contents of maleimide groups in polyamides.     

Cure monitoring via different techniques: Isothermal cure of an epoxy-novolac molding compound

The isothermal cure of an epoxy-novolac molding compound was studied by means of Fourier-transform infrared spectroscopy (FTIR) and dielectrometry (DE). Results obtained were compared with previous differential scanning calorimetric (DSC) observations. The behavior of epoxide conversion (_FTIR) measured via FTIR was found similar to (but not exactly coinciding with) the extent of cure (_DSC) determined previously by means of DSC. As for the DE analysis, directly measurable properties such as permittivity () and loss factor () varied in a complicated manner during the course of cure, showing strong dependence on both temperature and frequency. Other dielectric parameters (such as ionic conductivity, relaxed permittivity, and characteristic relaxation time) previously suggested in the literature as suitable for cure monitoring purposes were found difficult to determine within the limited frequency range (10⁰ to 10⁴ Hz) here. With some arbitrariness, the relative drop in log (at 100 Hz) was taken as an index (_DE) for the extent of cure. It was observed that _DE behaves in a manner similar to _FTIR and _DSC Comments on the application of these three techniques in the characterization of thermosetting systems were given.     

Synthesis and kinetic studies of UV‐curable urethane‐acrylates

The UV‐curable urethane‐acrylates based on 2‐hydroxyethyl methacrylate (HEMA)‐terminated polyurethane (PU) for lithographic and coating applications are investigated in this study. Series of PU prepolymers were made from 4,4‐diphenyl methane diisocyanate (MDI), poly(propylene oxide) glycol (PPG 400), poly(butylene adipate)glycol (PBA 500), or poly(tetramethylene oxide) glycol (PTMO 1000) and are terminated with HEMA. The 2,2‐azobisisobutyronitrile (AIBN) was used as a UV‐initiator under air atmosphere. The curing kinetics of HEMA‐terminated PU film were studied. The curing analysis, using FTIR and reaction kinetics, indicate the reaction rate equation correlates well with the film thickness [T], initiator concentration [I], unreacted double bond concentration [C?C], and exposed energy [E] of the reaction system. The kinetic rate equation for the UV‐curable reaction can be written as © 2004 Wiley Periodicals, Inc. J Appl Polym Sci 91: 3162–3166, 2004     

Morphology and characterization of conductive films based on polyaniline‐coated polystyrene latexes

The polyaniline (PANI)‐coated polystyrene (PS) latexes were synthesized, and the electrically conductive films were prepared thereafter. The weight ratio of PANI was 5%. Thermal analysis of the latices was performed using DSC and TGA. In this study, the electrically conductive films were prepared above the PS glass transition temperature (T_g). During the film formation, the effects of the annealing temperature and atmosphere (air or N₂) on the film resistance were investigated. In addition, the film morphology was observed utilizing scanning electron microscopy. The film resistance decreased in the initial heating stage due to the increasing temperature and the compaction of film. Then the film resistance increased with further annealing due to the aging of PANI. Typically, the film resistance was about 6000 Ω/sq, and the conductivity was 0.3 S/cm. © 2006 Wiley Periodicals, Inc. J Appl Polym Sci 102: 5406–5413, 2006     

Surface modification and physical properties of various UHMWPE‐fiber‐reinforced modified epoxy composites

Two surface modification methods—plasma surface treatment and chemical agent treatment—were used to investigate their effects on the surface properties of ultrahigh‐molecular‐weight polyethylene (UHMWPE) fibers. In the analyses, performed using electron spectroscopy for chemical analysis, changes in weight, and scanning electron microscope observations, demonstrated that the two fiber‐surface‐modified composites formed between UHMWPE fiber and epoxy matrix exhibited improved interfacial adhesion and slight improvements in tensile strengths, but notable decreases in elongation, relative to those properties of the composites reinforced with the untreated UHMWPE fibers. In addition, three kinds of epoxy resins—neat DGEBA, polyurethane‐crosslinked DGEBA, and BHHBP‐DGEBA—were used as resin matrices to examine the tensile and elongation properties of their UHMWPE fiber‐reinforced composites. From stress/strain measurements and scanning electron microscope observations, the resin matrix improved the tensile strength apparently, but did not affect the elongation. © 2007 Wiley Periodicals, Inc. J Appl Polym Sci 104: 655–665, 2007